Vectorize the CRC-32 calculation on the s390x.

Use vector extensions when compiling for s390x and binutils knows
about them. At runtime, check whether kernel supports vector
extensions (it has to be not just the CPU, but also the kernel)
and choose between the regular and the vectorized implementations.

Co-authored-by: Eduard Stefes <eddy@linux.ibm.com>

10 files changed, 439 insertions, 2 deletions

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 8aa8751f..8a8fde79 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -103,7 +103,7 @@ check_include_file(unistd.h HAVE_UNISTD_H)
 if(MSVC)
     set(CMAKE_REQUIRED_FLAGS "-WX")
 else(MSVC)
-    set(CMAKE_REQUIRED_FLAGS "-WError")
+    set(CMAKE_REQUIRED_FLAGS "-Werror")
 endif(MSVC)
 
 check_c_source_compiles(
diff --git a/Makefile.in b/Makefile.in
index 52d5c2c4..d9de5630 100644
--- a/Makefile.in
+++ b/Makefile.in
@@ -27,6 +27,7 @@ LDFLAGS=
 TEST_LIBS=-L. libz.a
 LDSHARED=$(CC)
 CPP=$(CC) -E
+VGFMAFLAG=
 
 STATICLIB=libz.a
 SHAREDLIB=libz.so
@@ -164,6 +165,9 @@ adler32.o: $(SRCDIR)adler32.c
 crc32.o: $(SRCDIR)crc32.c
         $(CC) $(CFLAGS) $(ZINC) -c -o $@ $(SRCDIR)crc32.c
 
+crc32_vx.o: $(SRCDIR)contrib/crc32vx/crc32_vx.c
+        $(CC) $(CFLAGS) $(VGFMAFLAG) $(ZINC) -c -o $@ $(SRCDIR)contrib/crc32vx/crc32_vx.c
+
 deflate.o: $(SRCDIR)deflate.c
         $(CC) $(CFLAGS) $(ZINC) -c -o $@ $(SRCDIR)deflate.c
 
@@ -214,6 +218,11 @@ crc32.lo: $(SRCDIR)crc32.c
         $(CC) $(SFLAGS) $(ZINC) -DPIC -c -o objs/crc32.o $(SRCDIR)crc32.c
         -@mv objs/crc32.o $@
 
+crc32_vx.lo: $(SRCDIR)contrib/crc32vx/crc32_vx.c
+        -@mkdir objs 2>/dev/null || test -d objs
+        $(CC) $(SFLAGS) $(VGFMAFLAG) $(ZINC) -DPIC -c -o objs/crc32_vx.o $(SRCDIR)contrib/crc32vx/crc32_vx.c
+        -@mv objs/crc32_vx.o $@
+
 deflate.lo: $(SRCDIR)deflate.c
         -@mkdir objs 2>/dev/null || test -d objs
         $(CC) $(SFLAGS) $(ZINC) -DPIC -c -o objs/deflate.o $(SRCDIR)deflate.c
@@ -406,6 +415,7 @@ infback.o inflate.o: $(SRCDIR)zutil.h $(SRCDIR)zlib.h zconf.h $(SRCDIR)inftrees.
 inffast.o: $(SRCDIR)zutil.h $(SRCDIR)zlib.h zconf.h $(SRCDIR)inftrees.h $(SRCDIR)inflate.h $(SRCDIR)inffast.h
 inftrees.o: $(SRCDIR)zutil.h $(SRCDIR)zlib.h zconf.h $(SRCDIR)inftrees.h
 trees.o: $(SRCDIR)deflate.h $(SRCDIR)zutil.h $(SRCDIR)zlib.h zconf.h $(SRCDIR)trees.h
+crc32_vx.o: $(SRCDIR)zutil.h $(SRCDIR)zlib.h zconf.h $(SRCDIR)contrib/crc32vx/crc32_vx_hooks.h
 
 adler32.lo: $(SRCDIR)zutil.h $(SRCDIR)zlib.h zconf.h
 zutil.lo: $(SRCDIR)zutil.h $(SRCDIR)zlib.h zconf.h $(SRCDIR)gzguts.h
@@ -417,3 +427,4 @@ infback.lo inflate.lo: $(SRCDIR)zutil.h $(SRCDIR)zlib.h zconf.h $(SRCDIR)inftree
 inffast.lo: $(SRCDIR)zutil.h $(SRCDIR)zlib.h zconf.h $(SRCDIR)inftrees.h $(SRCDIR)inflate.h $(SRCDIR)inffast.h
 inftrees.lo: $(SRCDIR)zutil.h $(SRCDIR)zlib.h zconf.h $(SRCDIR)inftrees.h
 trees.lo: $(SRCDIR)deflate.h $(SRCDIR)zutil.h $(SRCDIR)zlib.h zconf.h $(SRCDIR)trees.h
+crc32_vx.lo: $(SRCDIR)zutil.h $(SRCDIR)zlib.h zconf.h $(SRCDIR)contrib/crc32vx/crc32_vx_hooks.h
\ No newline at end of file
diff --git a/configure b/configure
index 3770b03d..05bee1f4 100755
--- a/configure
+++ b/configure
@@ -95,6 +95,7 @@ memory=0
 undefined=0
 insecure=0
 unknown=0
+enable_crcvx=1
 old_cc="$CC"
 old_cflags="$CFLAGS"
 OBJC='$(OBJZ) $(OBJG)'
@@ -122,6 +123,7 @@ case "$1" in
       echo '  configure [--const] [--zprefix] [--prefix=PREFIX]  [--eprefix=EXPREFIX]' | tee -a configure.log
       echo '    [--insecure] [--static] [--64] [--libdir=LIBDIR] [--sharedlibdir=LIBDIR]' | tee -a configure.log
       echo '    [--includedir=INCLUDEDIR] [--archs="-arch i386 -arch x86_64"]' | tee -a configure.log
+      echo '    [--disable-crcvx]' | tee -a configure.log
         exit 0 ;;
     -p*=* | --prefix=*) prefix=`echo $1 | sed 's/.*=//'`; shift ;;
     -e*=* | --eprefix=*) exec_prefix=`echo $1 | sed 's/.*=//'`; shift ;;
@@ -150,6 +152,7 @@ case "$1" in
     --memory) memory=1; shift ;;
     --undefined) undefined=1; shift ;;
     --insecure) insecure=1; shift ;;
+    --disable-crcvx) enable_crcvx=0; shift ;;
     *) unknown=1; echo "unknown option ignored: $1" | tee -a configure.log; shift;;
     esac
 done
@@ -888,6 +891,70 @@ EOF
   fi
 fi
 
+# check for ibm s390x build
+HAVE_S390X=0
+cat > $test.c << EOF
+#ifndef __s390x__
+    #error
+#endif
+EOF
+if try $CC -c $CFLAGS $test.c; then 
+    echo "Checking for s390x build ... Yes." | tee -a configure.log
+    HAVE_S390X=1
+else
+    echo "Checking for s390x build ... No." | tee -a configure.log
+fi
+
+# check for ibm s390x vx vector extensions
+HAVE_S390X_VX=0
+if test $HAVE_S390X -eq 1 && test $enable_crcvx -eq 1 ; then
+    # preset the compiler specific flags
+    if test $clang -eq 1; then
+        VGFMAFLAG=-fzvector
+    else
+        VGFMAFLAG=-mzarch
+    fi
+
+    cat > $test.c <<EOF
+#ifndef __s390x__
+#error
+#endif
+#include <vecintrin.h>
+int main(void) {
+    unsigned long long a __attribute__((vector_size(16))) = { 0 };
+    unsigned long long b __attribute__((vector_size(16))) = { 0 };
+    unsigned char c __attribute__((vector_size(16))) = { 0 };
+    c = vec_gfmsum_accum_128(a, b, c);
+    return c[0];
+}
+EOF
+
+    # cflags already contains a valid march
+    if try $CC -c $CFLAGS $VGFMAFLAG $test.c; then
+        echo "Checking for s390x vx vector extension ... Yes." | tee -a configure.log
+        HAVE_S390X_VX=1
+    # or set march for our compile units
+    elif try $CC -c $CFLAGS $VGFMAFLAG -march=z13 $test.c; then
+        echo "Checking for s390x vx vector extension (march=z13) ... Yes." | tee -a configure.log
+        HAVE_S390X_VX=1
+        VGFMAFLAG="$VGFMAFLAG -march=z13"
+    # else we are not on s390x
+    else 
+        echo "Checking for s390x vx vector extension ... No." | tee -a configure.log
+    fi
+
+    # prepare compiling for s390x
+    if test $HAVE_S390X_VX -eq 1; then
+        CFLAGS="$CFLAGS -DHAVE_S390X_VX"
+        SFLAGS="$SFLAGS -DHAVE_S390X_VX"
+        OBJC="$OBJC crc32_vx.o"
+        PIC_OBJC="$PIC_OBJC crc32_vx.lo"
+    else
+        # target has no vx extension
+        VGFMAFLAG=""
+    fi
+fi
+
 # show the results in the log
 echo >> configure.log
 echo ALL = $ALL >> configure.log
@@ -919,6 +986,9 @@ echo mandir = $mandir >> configure.log
 echo prefix = $prefix >> configure.log
 echo sharedlibdir = $sharedlibdir >> configure.log
 echo uname = $uname >> configure.log
+echo HAVE_S390X = $HAVE_S390X >> configure.log
+echo HAVE_S390X_VX = $HAVE_S390X_VX >> configure.log
+echo VGFMAFLAG = $VGFMAFLAG >> configure.log
 
 # update Makefile with the configure results
 sed < ${SRCDIR}Makefile.in "
diff --git a/contrib/CMakeLists.txt b/contrib/CMakeLists.txt
index 8ec3bbd9..37a3491f 100644
--- a/contrib/CMakeLists.txt
+++ b/contrib/CMakeLists.txt
@@ -24,9 +24,15 @@ function(zlib_add_contrib_lib name description dir)
 endfunction(zlib_add_contrib_lib name description dir)
 
 function(zlib_add_contrib_feature name description dir)
+    if(ARGC EQUAL 4)
+        set(default_on ${ARGV3})
+    else()
+        set(default_on Off)
+    endif()
+
     option(ZLIB_WITH_${name}
            "Enable build ${description}"
-           OFF)
+           ${default_on})
 
     if(ZLIB_WITH_${name})
         add_subdirectory(${dir}/)
@@ -38,6 +44,7 @@ zlib_add_contrib_feature("GVMAT64"
     gcc_gvmat64)
 
 zlib_add_contrib_feature(INFBACK9 "with support for method 9 deflate" infback9)
+zlib_add_contrib_feature(CRC32VX "with S390X-CRC32VX implementation" crc32vx On)
 zlib_add_contrib_lib(ADA "Ada bindings" ada)
 zlib_add_contrib_lib(BLAST "blast binary" blast)
 zlib_add_contrib_lib(IOSTREAM3 "IOStream C++ bindings V3" iostream3)
diff --git a/contrib/README.contrib b/contrib/README.contrib
index d9480eea..173f1d48 100644
--- a/contrib/README.contrib
+++ b/contrib/README.contrib
@@ -46,6 +46,9 @@ puff/       by Mark Adler <madler@alumni.caltech.edu>
         Small, low memory usage inflate.  Also serves to provide an
         unambiguous description of the deflate format.
 
+crc32vx/      by Ilya Leoshkevich <iii@linux.ibm.com>
+        Hardware-accelerated CRC32 on IBM Z with Z13 VX extension.
+
 testzlib/   by Gilles Vollant <info@winimage.com>
         Example of the use of zlib
 
diff --git a/contrib/crc32vx/CMakeLists.txt b/contrib/crc32vx/CMakeLists.txt
new file mode 100644
index 00000000..ee46fa3a
--- /dev/null
+++ b/contrib/crc32vx/CMakeLists.txt
@@ -0,0 +1,67 @@
+# check if we compile for IBM s390x
+#
+CHECK_C_SOURCE_COMPILES("
+#ifndef __s390x__
+    #error
+#endif
+int main() {return 0;}
+" HAS_S390X_SUPPORT)
+
+#
+# Check for IBM S390X - VX extensions
+#
+if(ZLIB_WITH_CRC32VX AND HAS_S390X_SUPPORT)
+    # preset the compiler specific flags
+    if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
+        set(VGFMAFLAG "-fzvector")
+    else()
+        set(VGFMAFLAG "-mzarch")
+    endif(CMAKE_C_COMPILER_ID STREQUAL "Clang")
+
+    set(S390X_VX_TEST
+        "#ifndef __s390x__ \n\
+        #error \n\
+        #endif \n\
+        #include <vecintrin.h> \n\
+        int main(void) { \
+           unsigned long long a __attribute__((vector_size(16))) = { 0 }; \
+           unsigned long long b __attribute__((vector_size(16))) = { 0 }; \
+           unsigned char c __attribute__((vector_size(16))) = { 0 }; \
+           c = vec_gfmsum_accum_128(a, b, c); \
+           return c[0]; \
+        }")
+
+    # cflags already contains a valid march
+    set(CMAKE_REQUIRED_FLAGS "${VGFMAFLAG}")
+    check_c_source_compiles("${S390X_VX_TEST}" HAS_S390X_VX_SUPPORT)
+    unset(CMAKE_REQUIRED_FLAGS)
+
+    # or set march for our compile units
+    if(NOT HAS_S390X_VX_SUPPORT)
+        set(CMAKE_REQUIRED_FLAGS "${VGFMAFLAG} -march=z13")
+        check_c_source_compiles("${S390X_VX_TEST}" HAS_Z13_S390X_VX_SUPPORT)
+        unset(CMAKE_REQUIRED_FLAGS )
+        list(APPEND VGFMAFLAG "-march=z13")
+    endif(NOT HAS_S390X_VX_SUPPORT)
+
+    # prepare compiling for s390x
+    if(HAS_S390X_VX_SUPPORT OR HAS_Z13_S390X_VX_SUPPORT)
+        if(ZLIB_BUILD_SHARED)
+            target_sources(zlib
+                PRIVATE
+                    crc32_vx.c
+                    crc32_vx_hooks.h)
+            target_compile_definitions(zlib PUBLIC -DHAVE_S390X_VX=1)
+        endif(ZLIB_BUILD_SHARED)
+        if(ZLIB_BUILD_STATIC)
+            target_sources(zlibstatic
+                PRIVATE
+                    crc32_vx.c
+                    crc32_vx_hooks.h)
+            target_compile_definitions(zlibstatic PUBLIC -DHAVE_S390X_VX=1)
+        endif(ZLIB_BUILD_STATIC)
+        set_source_files_properties(
+            crc32_vx.c
+            PROPERTIES COMPILE_OPTIONS "${VGFMAFLAG}")
+    endif(HAS_S390X_VX_SUPPORT OR HAS_Z13_S390X_VX_SUPPORT)
+endif(ZLIB_WITH_CRC32VX AND HAS_S390X_SUPPORT)
diff --git a/contrib/crc32vx/README b/contrib/crc32vx/README
new file mode 100644
index 00000000..329610d5
--- /dev/null
+++ b/contrib/crc32vx/README
@@ -0,0 +1,9 @@
+IBM Z mainframes starting from version z13 provide vector instructions, which 
+allows vectorization of crc32. This extension is build by default when targeting
+ibm s390x. However this extension can disabled if desired:
+
+    # for configure build
+    $ ./configure --disable-crcvx
+
+    # for cmake build
+    $ cmake .. -DZLIB_CRC32VX=off
diff --git a/contrib/crc32vx/crc32_vx.c b/contrib/crc32vx/crc32_vx.c
new file mode 100644
index 00000000..e718e340
--- /dev/null
+++ b/contrib/crc32vx/crc32_vx.c
@@ -0,0 +1,254 @@
+/*
+ * Hardware-accelerated CRC-32 variants for Linux on z Systems
+ *
+ * Use the z/Architecture Vector Extension Facility to accelerate the
+ * computing of bitreflected CRC-32 checksums.
+ *
+ * This CRC-32 implementation algorithm is bitreflected and processes
+ * the least-significant bit first (Little-Endian).
+ *
+ * This code was originally written by Hendrik Brueckner
+ * <brueckner@linux.vnet.ibm.com> for use in the Linux kernel and has been
+ * relicensed under the zlib license.
+ */
+#define Z_ONCE
+#include "../../zutil.h"
+#include "crc32_vx_hooks.h"
+
+#include <stdint.h>
+#include <stdio.h>
+#include <vecintrin.h>
+#include <sys/auxv.h>
+
+#ifdef __clang__
+#  if ((__clang_major__ == 18) || (__clang_major__ == 19 && (__clang_minor__ < 1 || (__clang_minor__ == 1 && __clang_patchlevel__ < 2))))
+# error crc32_vx optimizations are broken due to compiler bug in Clang versions: 18.0.0 <= clang_version < 19.1.2. \
+        Either disable the zlib crc32_vx optimization, or switch to another compiler/compiler version.
+#  endif
+#endif
+
+#define VX_MIN_LEN 64
+#define VX_ALIGNMENT 16L
+#define VX_ALIGN_MASK (VX_ALIGNMENT - 1)
+
+typedef unsigned char uv16qi __attribute__((vector_size(16)));
+typedef unsigned int uv4si __attribute__((vector_size(16)));
+typedef unsigned long long uv2di __attribute__((vector_size(16)));
+
+local uint32_t crc32_le_vgfm_16(uint32_t crc, const unsigned char *buf, size_t len) {
+    /*
+     * The CRC-32 constant block contains reduction constants to fold and
+     * process particular chunks of the input data stream in parallel.
+     *
+     * For the CRC-32 variants, the constants are precomputed according to
+     * these definitions:
+     *
+     *      R1 = [(x4*128+32 mod P'(x) << 32)]' << 1
+     *      R2 = [(x4*128-32 mod P'(x) << 32)]' << 1
+     *      R3 = [(x128+32 mod P'(x) << 32)]'   << 1
+     *      R4 = [(x128-32 mod P'(x) << 32)]'   << 1
+     *      R5 = [(x64 mod P'(x) << 32)]'       << 1
+     *      R6 = [(x32 mod P'(x) << 32)]'       << 1
+     *
+     *      The bitreflected Barret reduction constant, u', is defined as
+     *      the bit reversal of floor(x**64 / P(x)).
+     *
+     *      where P(x) is the polynomial in the normal domain and the P'(x) is the
+     *      polynomial in the reversed (bitreflected) domain.
+     *
+     * CRC-32 (IEEE 802.3 Ethernet, ...) polynomials:
+     *
+     *      P(x)  = 0x04C11DB7
+     *      P'(x) = 0xEDB88320
+     */
+    const uv16qi perm_le2be = {15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0};  /* BE->LE mask */
+    const uv2di r2r1 = {0x1C6E41596, 0x154442BD4};                                     /* R2, R1 */
+    const uv2di r4r3 = {0x0CCAA009E, 0x1751997D0};                                     /* R4, R3 */
+    const uv2di r5 = {0, 0x163CD6124};                                                 /* R5 */
+    const uv2di ru_poly = {0, 0x1F7011641};                                            /* u' */
+    const uv2di crc_poly = {0, 0x1DB710641};                                           /* P'(x) << 1 */
+
+    /*
+     * Load the initial CRC value.
+     *
+     * The CRC value is loaded into the rightmost word of the
+     * vector register and is later XORed with the LSB portion
+     * of the loaded input data.
+     */
+    uv2di v0 = {0, 0};
+    v0 = (uv2di)vec_insert(crc, (uv4si)v0, 3);
+
+    /* Load a 64-byte data chunk and XOR with CRC */
+    uv2di v1 = vec_perm(((uv2di *)buf)[0], ((uv2di *)buf)[0], perm_le2be);
+    uv2di v2 = vec_perm(((uv2di *)buf)[1], ((uv2di *)buf)[1], perm_le2be);
+    uv2di v3 = vec_perm(((uv2di *)buf)[2], ((uv2di *)buf)[2], perm_le2be);
+    uv2di v4 = vec_perm(((uv2di *)buf)[3], ((uv2di *)buf)[3], perm_le2be);
+
+    v1 ^= v0;
+    buf += 64;
+    len -= 64;
+
+    while (len >= 64) {
+        /* Load the next 64-byte data chunk */
+        uv16qi part1 = vec_perm(((uv16qi *)buf)[0], ((uv16qi *)buf)[0], perm_le2be);
+        uv16qi part2 = vec_perm(((uv16qi *)buf)[1], ((uv16qi *)buf)[1], perm_le2be);
+        uv16qi part3 = vec_perm(((uv16qi *)buf)[2], ((uv16qi *)buf)[2], perm_le2be);
+        uv16qi part4 = vec_perm(((uv16qi *)buf)[3], ((uv16qi *)buf)[3], perm_le2be);
+
+        /*
+         * Perform a GF(2) multiplication of the doublewords in V1 with
+         * the R1 and R2 reduction constants in V0.  The intermediate result
+         * is then folded (accumulated) with the next data chunk in PART1 and
+         * stored in V1. Repeat this step for the register contents
+         * in V2, V3, and V4 respectively.
+         */
+        v1 = (uv2di)vec_gfmsum_accum_128(r2r1, v1, part1);
+        v2 = (uv2di)vec_gfmsum_accum_128(r2r1, v2, part2);
+        v3 = (uv2di)vec_gfmsum_accum_128(r2r1, v3, part3);
+        v4 = (uv2di)vec_gfmsum_accum_128(r2r1, v4, part4);
+
+        buf += 64;
+        len -= 64;
+    }
+
+    /*
+     * Fold V1 to V4 into a single 128-bit value in V1.  Multiply V1 with R3
+     * and R4 and accumulating the next 128-bit chunk until a single 128-bit
+     * value remains.
+     */
+    v1 = (uv2di)vec_gfmsum_accum_128(r4r3, v1, (uv16qi)v2);
+    v1 = (uv2di)vec_gfmsum_accum_128(r4r3, v1, (uv16qi)v3);
+    v1 = (uv2di)vec_gfmsum_accum_128(r4r3, v1, (uv16qi)v4);
+
+    while (len >= 16) {
+        /* Load next data chunk */
+        v2 = vec_perm(*(uv2di *)buf, *(uv2di *)buf, perm_le2be);
+
+        /* Fold next data chunk */
+        v1 = (uv2di)vec_gfmsum_accum_128(r4r3, v1, (uv16qi)v2);
+
+        buf += 16;
+        len -= 16;
+    }
+
+    /*
+     * Set up a vector register for byte shifts.  The shift value must
+     * be loaded in bits 1-4 in byte element 7 of a vector register.
+     * Shift by 8 bytes: 0x40
+     * Shift by 4 bytes: 0x20
+     */
+    uv16qi v9 = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
+    v9 = vec_insert((unsigned char)0x40, v9, 7);
+
+    /*
+     * Prepare V0 for the next GF(2) multiplication: shift V0 by 8 bytes
+     * to move R4 into the rightmost doubleword and set the leftmost
+     * doubleword to 0x1.
+     */
+    v0 = vec_srb(r4r3, (uv2di)v9);
+    v0[0] = 1;
+
+    /*
+     * Compute GF(2) product of V1 and V0.  The rightmost doubleword
+     * of V1 is multiplied with R4.  The leftmost doubleword of V1 is
+     * multiplied by 0x1 and is then XORed with rightmost product.
+     * Implicitly, the intermediate leftmost product becomes padded
+     */
+    v1 = (uv2di)vec_gfmsum_128(v0, v1);
+
+    /*
+     * Now do the final 32-bit fold by multiplying the rightmost word
+     * in V1 with R5 and XOR the result with the remaining bits in V1.
+     *
+     * To achieve this by a single VGFMAG, right shift V1 by a word
+     * and store the result in V2 which is then accumulated.  Use the
+     * vector unpack instruction to load the rightmost half of the
+     * doubleword into the rightmost doubleword element of V1; the other
+     * half is loaded in the leftmost doubleword.
+     * The vector register with CONST_R5 contains the R5 constant in the
+     * rightmost doubleword and the leftmost doubleword is zero to ignore
+     * the leftmost product of V1.
+     */
+    v9 = vec_insert((unsigned char)0x20, v9, 7);
+    v2 = vec_srb(v1, (uv2di)v9);
+    v1 = vec_unpackl((uv4si)v1);  /* Split rightmost doubleword */
+    v1 = (uv2di)vec_gfmsum_accum_128(r5, v1, (uv16qi)v2);
+
+    /*
+     * Apply a Barret reduction to compute the final 32-bit CRC value.
+     *
+     * The input values to the Barret reduction are the degree-63 polynomial
+     * in V1 (R(x)), degree-32 generator polynomial, and the reduction
+     * constant u.  The Barret reduction result is the CRC value of R(x) mod
+     * P(x).
+     *
+     * The Barret reduction algorithm is defined as:
+     *
+     *    1. T1(x) = floor( R(x) / x^32 ) GF2MUL u
+     *    2. T2(x) = floor( T1(x) / x^32 ) GF2MUL P(x)
+     *    3. C(x)  = R(x) XOR T2(x) mod x^32
+     *
+     *  Note: The leftmost doubleword of vector register containing
+     *  CONST_RU_POLY is zero and, thus, the intermediate GF(2) product
+     *  is zero and does not contribute to the final result.
+     */
+
+    /* T1(x) = floor( R(x) / x^32 ) GF2MUL u */
+    v2 = vec_unpackl((uv4si)v1);
+    v2 = (uv2di)vec_gfmsum_128(ru_poly, v2);
+
+    /*
+     * Compute the GF(2) product of the CRC polynomial with T1(x) in
+     * V2 and XOR the intermediate result, T2(x), with the value in V1.
+     * The final result is stored in word element 2 of V2.
+     */
+    v2 = vec_unpackl((uv4si)v2);
+    v2 = (uv2di)vec_gfmsum_accum_128(crc_poly, v2, (uv16qi)v1);
+
+    return ((uv4si)v2)[2];
+}
+
+
+local unsigned long s390_crc32_vx(unsigned long crc, const unsigned char FAR *buf, z_size_t len)
+{
+    uintptr_t prealign, aligned, remaining;
+
+    if (buf == Z_NULL) return 0UL;
+
+    if (len < VX_MIN_LEN + VX_ALIGN_MASK)
+        return crc32_z(crc, buf, len);
+
+    if ((uintptr_t)buf & VX_ALIGN_MASK) {
+        prealign = VX_ALIGNMENT - ((uintptr_t)buf & VX_ALIGN_MASK);
+        len -= prealign;
+        crc = crc32_z(crc, buf, prealign);
+        buf += prealign;
+    }
+    aligned = len & ~VX_ALIGN_MASK;
+    remaining = len & VX_ALIGN_MASK;
+
+    crc = crc32_le_vgfm_16(crc ^ 0xffffffff, buf, (size_t)aligned) ^ 0xffffffff;
+
+    if (remaining)
+        crc = crc32_z(crc, buf + aligned, remaining);
+
+    return crc;
+}
+
+local z_once_t s390_crc32_made = Z_ONCE_INIT;
+local void s390_crc32_setup() {
+    unsigned long hwcap = getauxval(AT_HWCAP);
+
+    if (hwcap & HWCAP_S390_VX)
+        crc32_z_hook = s390_crc32_vx;
+    else
+        crc32_z_hook = crc32_z;
+}
+
+local unsigned long s390_crc32_init(unsigned long crc, const unsigned char FAR *buf, z_size_t len)
+{
+    z_once(&s390_crc32_made,s390_crc32_setup);
+    return crc32_z_hook(crc, buf, len);
+}
+
+ZLIB_INTERNAL unsigned long (*crc32_z_hook)(unsigned long crc, const unsigned char FAR *buf, z_size_t len) = s390_crc32_init;
diff --git a/contrib/crc32vx/crc32_vx_hooks.h b/contrib/crc32vx/crc32_vx_hooks.h
new file mode 100644
index 00000000..951c3188
--- /dev/null
+++ b/contrib/crc32vx/crc32_vx_hooks.h
@@ -0,0 +1,9 @@
+#ifndef CRC32_VX_HOOKS_H
+#define CRC32_VX_HOOKS_H
+
+/**
+ * CRC HOOKS
+ */
+ZLIB_INTERNAL extern unsigned long (*crc32_z_hook)(unsigned long crc, const unsigned char FAR *buf, z_size_t len);
+
+#endif /* CRC32_VX_HOOKS_H */
diff --git a/crc32.c b/crc32.c
index 4cc573f8..4d5f5b23 100644
--- a/crc32.c
+++ b/crc32.c
@@ -32,6 +32,10 @@
 
 #include "zutil.h"      /* for Z_U4, Z_U8, z_crc_t, and FAR definitions */
 
+#ifdef HAVE_S390X_VX
+#  include "contrib/crc32vx/crc32_vx_hooks.h"
+#endif
+
  /*
   A CRC of a message is computed on N braids of words in the message, where
   each word consists of W bytes (4 or 8). If N is 3, for example, then three
@@ -942,6 +946,9 @@ unsigned long ZEXPORT crc32_z(unsigned long crc, const unsigned char FAR *buf,
 /* ========================================================================= */
 unsigned long ZEXPORT crc32(unsigned long crc, const unsigned char FAR *buf,
                             uInt len) {
+    #ifdef HAVE_S390X_VX
+    return crc32_z_hook(crc, buf, len);
+    #endif
     return crc32_z(crc, buf, len);
 }