xref: /aosp_15_r20/external/zlib/patches/0001-simd.patch (revision 86ee64e75fa5f8bce2c8c356138035642429cd05)

diff --git a/crc32.c b/crc32.c
index 9580440c0e6b..9162429cc7b4 100644
--- a/crc32.c
+++ b/crc32.c
@@ -28,6 +28,8 @@
 #  endif /* !DYNAMIC_CRC_TABLE */
 #endif /* MAKECRCH */

+#include "deflate.h"
+#include "x86.h"
 #include "zutil.h"      /* for STDC and FAR definitions */

 /* Definitions for doing the crc four data bytes at a time. */
@@ -440,3 +442,28 @@ uLong ZEXPORT crc32_combine64(crc1, crc2, len2)
 {
     return crc32_combine_(crc1, crc2, len2);
 }
+
+ZLIB_INTERNAL void crc_reset(deflate_state *const s)
+{
+    if (x86_cpu_enable_simd) {
+        crc_fold_init(s);
+        return;
+    }
+    s->strm->adler = crc32(0L, Z_NULL, 0);
+}
+
+ZLIB_INTERNAL void crc_finalize(deflate_state *const s)
+{
+    if (x86_cpu_enable_simd)
+        s->strm->adler = crc_fold_512to32(s);
+}
+
+ZLIB_INTERNAL void copy_with_crc(z_streamp strm, Bytef *dst, long size)
+{
+    if (x86_cpu_enable_simd) {
+        crc_fold_copy(strm->state, dst, strm->next_in, size);
+        return;
+    }
+    zmemcpy(dst, strm->next_in, size);
+    strm->adler = crc32(strm->adler, dst, size);
+}
diff --git a/crc_folding.c b/crc_folding.c
new file mode 100644
index 000000000000..48d77744aaf4
--- /dev/null
+++ b/crc_folding.c
@@ -0,0 +1,493 @@
+/*
+ * Compute the CRC32 using a parallelized folding approach with the PCLMULQDQ
+ * instruction.
+ *
+ * A white paper describing this algorithm can be found at:
+ * http://www.intel.com/content/dam/www/public/us/en/documents/white-papers/fast-crc-computation-generic-polynomials-pclmulqdq-paper.pdf
+ *
+ * Copyright (C) 2013 Intel Corporation. All rights reserved.
+ * Authors:
+ * 	Wajdi Feghali   <[email protected]>
+ * 	Jim Guilford    <[email protected]>
+ * 	Vinodh Gopal    <[email protected]>
+ * 	Erdinc Ozturk   <[email protected]>
+ * 	Jim Kukunas     <[email protected]>
+ *
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+#include "deflate.h"
+
+#include <inttypes.h>
+#include <emmintrin.h>
+#include <immintrin.h>
+#include <wmmintrin.h>
+
+#define CRC_LOAD(s) \
+    do { \
+        __m128i xmm_crc0 = _mm_loadu_si128((__m128i *)s->crc0 + 0);\
+        __m128i xmm_crc1 = _mm_loadu_si128((__m128i *)s->crc0 + 1);\
+        __m128i xmm_crc2 = _mm_loadu_si128((__m128i *)s->crc0 + 2);\
+        __m128i xmm_crc3 = _mm_loadu_si128((__m128i *)s->crc0 + 3);\
+        __m128i xmm_crc_part = _mm_loadu_si128((__m128i *)s->crc0 + 4);
+
+#define CRC_SAVE(s) \
+        _mm_storeu_si128((__m128i *)s->crc0 + 0, xmm_crc0);\
+        _mm_storeu_si128((__m128i *)s->crc0 + 1, xmm_crc1);\
+        _mm_storeu_si128((__m128i *)s->crc0 + 2, xmm_crc2);\
+        _mm_storeu_si128((__m128i *)s->crc0 + 3, xmm_crc3);\
+        _mm_storeu_si128((__m128i *)s->crc0 + 4, xmm_crc_part);\
+    } while (0);
+
+ZLIB_INTERNAL void crc_fold_init(deflate_state *const s)
+{
+    CRC_LOAD(s)
+
+    xmm_crc0 = _mm_cvtsi32_si128(0x9db42487);
+    xmm_crc1 = _mm_setzero_si128();
+    xmm_crc2 = _mm_setzero_si128();
+    xmm_crc3 = _mm_setzero_si128();
+
+    CRC_SAVE(s)
+
+    s->strm->adler = 0;
+}
+
+local void fold_1(deflate_state *const s,
+        __m128i *xmm_crc0, __m128i *xmm_crc1,
+        __m128i *xmm_crc2, __m128i *xmm_crc3)
+{
+    const __m128i xmm_fold4 = _mm_set_epi32(
+            0x00000001, 0x54442bd4,
+            0x00000001, 0xc6e41596);
+
+    __m128i x_tmp3;
+    __m128 ps_crc0, ps_crc3, ps_res;
+
+    x_tmp3 = *xmm_crc3;
+
+    *xmm_crc3 = *xmm_crc0;
+    *xmm_crc0 = _mm_clmulepi64_si128(*xmm_crc0, xmm_fold4, 0x01);
+    *xmm_crc3 = _mm_clmulepi64_si128(*xmm_crc3, xmm_fold4, 0x10);
+    ps_crc0 = _mm_castsi128_ps(*xmm_crc0);
+    ps_crc3 = _mm_castsi128_ps(*xmm_crc3);
+    ps_res = _mm_xor_ps(ps_crc0, ps_crc3);
+
+    *xmm_crc0 = *xmm_crc1;
+    *xmm_crc1 = *xmm_crc2;
+    *xmm_crc2 = x_tmp3;
+    *xmm_crc3 = _mm_castps_si128(ps_res);
+}
+
+local void fold_2(deflate_state *const s,
+        __m128i *xmm_crc0, __m128i *xmm_crc1,
+        __m128i *xmm_crc2, __m128i *xmm_crc3)
+{
+    const __m128i xmm_fold4 = _mm_set_epi32(
+            0x00000001, 0x54442bd4,
+            0x00000001, 0xc6e41596);
+
+    __m128i x_tmp3, x_tmp2;
+    __m128 ps_crc0, ps_crc1, ps_crc2, ps_crc3, ps_res31, ps_res20;
+
+    x_tmp3 = *xmm_crc3;
+    x_tmp2 = *xmm_crc2;
+
+    *xmm_crc3 = *xmm_crc1;
+    *xmm_crc1 = _mm_clmulepi64_si128(*xmm_crc1, xmm_fold4, 0x01);
+    *xmm_crc3 = _mm_clmulepi64_si128(*xmm_crc3, xmm_fold4, 0x10);
+    ps_crc3 = _mm_castsi128_ps(*xmm_crc3);
+    ps_crc1 = _mm_castsi128_ps(*xmm_crc1);
+    ps_res31= _mm_xor_ps(ps_crc3, ps_crc1);
+
+    *xmm_crc2 = *xmm_crc0;
+    *xmm_crc0 = _mm_clmulepi64_si128(*xmm_crc0, xmm_fold4, 0x01);
+    *xmm_crc2 = _mm_clmulepi64_si128(*xmm_crc2, xmm_fold4, 0x10);
+    ps_crc0 = _mm_castsi128_ps(*xmm_crc0);
+    ps_crc2 = _mm_castsi128_ps(*xmm_crc2);
+    ps_res20= _mm_xor_ps(ps_crc0, ps_crc2);
+
+    *xmm_crc0 = x_tmp2;
+    *xmm_crc1 = x_tmp3;
+    *xmm_crc2 = _mm_castps_si128(ps_res20);
+    *xmm_crc3 = _mm_castps_si128(ps_res31);
+}
+
+local void fold_3(deflate_state *const s,
+        __m128i *xmm_crc0, __m128i *xmm_crc1,
+        __m128i *xmm_crc2, __m128i *xmm_crc3)
+{
+    const __m128i xmm_fold4 = _mm_set_epi32(
+            0x00000001, 0x54442bd4,
+            0x00000001, 0xc6e41596);
+
+    __m128i x_tmp3;
+    __m128 ps_crc0, ps_crc1, ps_crc2, ps_crc3, ps_res32, ps_res21, ps_res10;
+
+    x_tmp3 = *xmm_crc3;
+
+    *xmm_crc3 = *xmm_crc2;
+    *xmm_crc2 = _mm_clmulepi64_si128(*xmm_crc2, xmm_fold4, 0x01);
+    *xmm_crc3 = _mm_clmulepi64_si128(*xmm_crc3, xmm_fold4, 0x10);
+    ps_crc2 = _mm_castsi128_ps(*xmm_crc2);
+    ps_crc3 = _mm_castsi128_ps(*xmm_crc3);
+    ps_res32 = _mm_xor_ps(ps_crc2, ps_crc3);
+
+    *xmm_crc2 = *xmm_crc1;
+    *xmm_crc1 = _mm_clmulepi64_si128(*xmm_crc1, xmm_fold4, 0x01);
+    *xmm_crc2 = _mm_clmulepi64_si128(*xmm_crc2, xmm_fold4, 0x10);
+    ps_crc1 = _mm_castsi128_ps(*xmm_crc1);
+    ps_crc2 = _mm_castsi128_ps(*xmm_crc2);
+    ps_res21= _mm_xor_ps(ps_crc1, ps_crc2);
+
+    *xmm_crc1 = *xmm_crc0;
+    *xmm_crc0 = _mm_clmulepi64_si128(*xmm_crc0, xmm_fold4, 0x01);
+    *xmm_crc1 = _mm_clmulepi64_si128(*xmm_crc1, xmm_fold4, 0x10);
+    ps_crc0 = _mm_castsi128_ps(*xmm_crc0);
+    ps_crc1 = _mm_castsi128_ps(*xmm_crc1);
+    ps_res10= _mm_xor_ps(ps_crc0, ps_crc1);
+
+    *xmm_crc0 = x_tmp3;
+    *xmm_crc1 = _mm_castps_si128(ps_res10);
+    *xmm_crc2 = _mm_castps_si128(ps_res21);
+    *xmm_crc3 = _mm_castps_si128(ps_res32);
+}
+
+local void fold_4(deflate_state *const s,
+        __m128i *xmm_crc0, __m128i *xmm_crc1,
+        __m128i *xmm_crc2, __m128i *xmm_crc3)
+{
+    const __m128i xmm_fold4 = _mm_set_epi32(
+            0x00000001, 0x54442bd4,
+            0x00000001, 0xc6e41596);
+
+    __m128i x_tmp0, x_tmp1, x_tmp2, x_tmp3;
+    __m128 ps_crc0, ps_crc1, ps_crc2, ps_crc3;
+    __m128 ps_t0, ps_t1, ps_t2, ps_t3;
+    __m128 ps_res0, ps_res1, ps_res2, ps_res3;
+
+    x_tmp0 = *xmm_crc0;
+    x_tmp1 = *xmm_crc1;
+    x_tmp2 = *xmm_crc2;
+    x_tmp3 = *xmm_crc3;
+
+    *xmm_crc0 = _mm_clmulepi64_si128(*xmm_crc0, xmm_fold4, 0x01);
+    x_tmp0 = _mm_clmulepi64_si128(x_tmp0, xmm_fold4, 0x10);
+    ps_crc0 = _mm_castsi128_ps(*xmm_crc0);
+    ps_t0 = _mm_castsi128_ps(x_tmp0);
+    ps_res0 = _mm_xor_ps(ps_crc0, ps_t0);
+
+    *xmm_crc1 = _mm_clmulepi64_si128(*xmm_crc1, xmm_fold4, 0x01);
+    x_tmp1 = _mm_clmulepi64_si128(x_tmp1, xmm_fold4, 0x10);
+    ps_crc1 = _mm_castsi128_ps(*xmm_crc1);
+    ps_t1 = _mm_castsi128_ps(x_tmp1);
+    ps_res1 = _mm_xor_ps(ps_crc1, ps_t1);
+
+    *xmm_crc2 = _mm_clmulepi64_si128(*xmm_crc2, xmm_fold4, 0x01);
+    x_tmp2 = _mm_clmulepi64_si128(x_tmp2, xmm_fold4, 0x10);
+    ps_crc2 = _mm_castsi128_ps(*xmm_crc2);
+    ps_t2 = _mm_castsi128_ps(x_tmp2);
+    ps_res2 = _mm_xor_ps(ps_crc2, ps_t2);
+
+    *xmm_crc3 = _mm_clmulepi64_si128(*xmm_crc3, xmm_fold4, 0x01);
+    x_tmp3 = _mm_clmulepi64_si128(x_tmp3, xmm_fold4, 0x10);
+    ps_crc3 = _mm_castsi128_ps(*xmm_crc3);
+    ps_t3 = _mm_castsi128_ps(x_tmp3);
+    ps_res3 = _mm_xor_ps(ps_crc3, ps_t3);
+
+    *xmm_crc0 = _mm_castps_si128(ps_res0);
+    *xmm_crc1 = _mm_castps_si128(ps_res1);
+    *xmm_crc2 = _mm_castps_si128(ps_res2);
+    *xmm_crc3 = _mm_castps_si128(ps_res3);
+}
+
+local const unsigned zalign(32) pshufb_shf_table[60] = {
+	0x84838281,0x88878685,0x8c8b8a89,0x008f8e8d, /* shl 15 (16 - 1)/shr1 */
+	0x85848382,0x89888786,0x8d8c8b8a,0x01008f8e, /* shl 14 (16 - 3)/shr2 */
+	0x86858483,0x8a898887,0x8e8d8c8b,0x0201008f, /* shl 13 (16 - 4)/shr3 */
+	0x87868584,0x8b8a8988,0x8f8e8d8c,0x03020100, /* shl 12 (16 - 4)/shr4 */
+	0x88878685,0x8c8b8a89,0x008f8e8d,0x04030201, /* shl 11 (16 - 5)/shr5 */
+	0x89888786,0x8d8c8b8a,0x01008f8e,0x05040302, /* shl 10 (16 - 6)/shr6 */
+	0x8a898887,0x8e8d8c8b,0x0201008f,0x06050403, /* shl  9 (16 - 7)/shr7 */
+	0x8b8a8988,0x8f8e8d8c,0x03020100,0x07060504, /* shl  8 (16 - 8)/shr8 */
+	0x8c8b8a89,0x008f8e8d,0x04030201,0x08070605, /* shl  7 (16 - 9)/shr9 */
+	0x8d8c8b8a,0x01008f8e,0x05040302,0x09080706, /* shl  6 (16 -10)/shr10*/
+	0x8e8d8c8b,0x0201008f,0x06050403,0x0a090807, /* shl  5 (16 -11)/shr11*/
+	0x8f8e8d8c,0x03020100,0x07060504,0x0b0a0908, /* shl  4 (16 -12)/shr12*/
+	0x008f8e8d,0x04030201,0x08070605,0x0c0b0a09, /* shl  3 (16 -13)/shr13*/
+	0x01008f8e,0x05040302,0x09080706,0x0d0c0b0a, /* shl  2 (16 -14)/shr14*/
+	0x0201008f,0x06050403,0x0a090807,0x0e0d0c0b  /* shl  1 (16 -15)/shr15*/
+};
+
+local void partial_fold(deflate_state *const s, const size_t len,
+        __m128i *xmm_crc0, __m128i *xmm_crc1,
+        __m128i *xmm_crc2, __m128i *xmm_crc3,
+        __m128i *xmm_crc_part)
+{
+
+    const __m128i xmm_fold4 = _mm_set_epi32(
+            0x00000001, 0x54442bd4,
+            0x00000001, 0xc6e41596);
+    const __m128i xmm_mask3 = _mm_set1_epi32(0x80808080);
+
+    __m128i xmm_shl, xmm_shr, xmm_tmp1, xmm_tmp2, xmm_tmp3;
+    __m128i xmm_a0_0, xmm_a0_1;
+    __m128 ps_crc3, psa0_0, psa0_1, ps_res;
+
+    xmm_shl = _mm_load_si128((__m128i *)pshufb_shf_table + (len - 1));
+    xmm_shr = xmm_shl;
+    xmm_shr = _mm_xor_si128(xmm_shr, xmm_mask3);
+
+    xmm_a0_0 = _mm_shuffle_epi8(*xmm_crc0, xmm_shl);
+
+    *xmm_crc0 = _mm_shuffle_epi8(*xmm_crc0, xmm_shr);
+    xmm_tmp1 = _mm_shuffle_epi8(*xmm_crc1, xmm_shl);
+    *xmm_crc0 = _mm_or_si128(*xmm_crc0, xmm_tmp1);
+
+    *xmm_crc1 = _mm_shuffle_epi8(*xmm_crc1, xmm_shr);
+    xmm_tmp2 = _mm_shuffle_epi8(*xmm_crc2, xmm_shl);
+    *xmm_crc1 = _mm_or_si128(*xmm_crc1, xmm_tmp2);
+
+    *xmm_crc2 = _mm_shuffle_epi8(*xmm_crc2, xmm_shr);
+    xmm_tmp3 = _mm_shuffle_epi8(*xmm_crc3, xmm_shl);
+    *xmm_crc2 = _mm_or_si128(*xmm_crc2, xmm_tmp3);
+
+    *xmm_crc3 = _mm_shuffle_epi8(*xmm_crc3, xmm_shr);
+    *xmm_crc_part = _mm_shuffle_epi8(*xmm_crc_part, xmm_shl);
+    *xmm_crc3 = _mm_or_si128(*xmm_crc3, *xmm_crc_part);
+
+    xmm_a0_1 = _mm_clmulepi64_si128(xmm_a0_0, xmm_fold4, 0x10);
+    xmm_a0_0 = _mm_clmulepi64_si128(xmm_a0_0, xmm_fold4, 0x01);
+
+    ps_crc3 = _mm_castsi128_ps(*xmm_crc3);
+    psa0_0 = _mm_castsi128_ps(xmm_a0_0);
+    psa0_1 = _mm_castsi128_ps(xmm_a0_1);
+
+    ps_res = _mm_xor_ps(ps_crc3, psa0_0);
+    ps_res = _mm_xor_ps(ps_res, psa0_1);
+
+    *xmm_crc3 = _mm_castps_si128(ps_res);
+}
+
+ZLIB_INTERNAL void crc_fold_copy(deflate_state *const s,
+        unsigned char *dst, const unsigned char *src, long len)
+{
+    unsigned long algn_diff;
+    __m128i xmm_t0, xmm_t1, xmm_t2, xmm_t3;
+
+    CRC_LOAD(s)
+
+    if (len < 16) {
+        if (len == 0)
+            return;
+        goto partial;
+    }
+
+    algn_diff = 0 - (uintptr_t)src & 0xF;
+    if (algn_diff) {
+        xmm_crc_part = _mm_loadu_si128((__m128i *)src);
+        _mm_storeu_si128((__m128i *)dst, xmm_crc_part);
+
+        dst += algn_diff;
+        src += algn_diff;
+        len -= algn_diff;
+
+        partial_fold(s, algn_diff, &xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3,
+            &xmm_crc_part);
+    }
+
+    while ((len -= 64) >= 0) {
+        xmm_t0 = _mm_load_si128((__m128i *)src);
+        xmm_t1 = _mm_load_si128((__m128i *)src + 1);
+        xmm_t2 = _mm_load_si128((__m128i *)src + 2);
+        xmm_t3 = _mm_load_si128((__m128i *)src + 3);
+
+        fold_4(s, &xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3);
+
+        _mm_storeu_si128((__m128i *)dst, xmm_t0);
+        _mm_storeu_si128((__m128i *)dst + 1, xmm_t1);
+        _mm_storeu_si128((__m128i *)dst + 2, xmm_t2);
+        _mm_storeu_si128((__m128i *)dst + 3, xmm_t3);
+
+        xmm_crc0 = _mm_xor_si128(xmm_crc0, xmm_t0);
+        xmm_crc1 = _mm_xor_si128(xmm_crc1, xmm_t1);
+        xmm_crc2 = _mm_xor_si128(xmm_crc2, xmm_t2);
+        xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_t3);
+
+        src += 64;
+        dst += 64;
+    }
+
+    /*
+     * len = num bytes left - 64
+     */
+    if (len + 16 >= 0) {
+        len += 16;
+
+        xmm_t0 = _mm_load_si128((__m128i *)src);
+        xmm_t1 = _mm_load_si128((__m128i *)src + 1);
+        xmm_t2 = _mm_load_si128((__m128i *)src + 2);
+
+        fold_3(s, &xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3);
+
+        _mm_storeu_si128((__m128i *)dst, xmm_t0);
+        _mm_storeu_si128((__m128i *)dst + 1, xmm_t1);
+        _mm_storeu_si128((__m128i *)dst + 2, xmm_t2);
+
+        xmm_crc1 = _mm_xor_si128(xmm_crc1, xmm_t0);
+        xmm_crc2 = _mm_xor_si128(xmm_crc2, xmm_t1);
+        xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_t2);
+
+        if (len == 0)
+            goto done;
+
+        dst += 48;
+        src += 48;
+    } else if (len + 32 >= 0) {
+        len += 32;
+
+        xmm_t0 = _mm_load_si128((__m128i *)src);
+        xmm_t1 = _mm_load_si128((__m128i *)src + 1);
+
+        fold_2(s, &xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3);
+
+        _mm_storeu_si128((__m128i *)dst, xmm_t0);
+        _mm_storeu_si128((__m128i *)dst + 1, xmm_t1);
+
+        xmm_crc2 = _mm_xor_si128(xmm_crc2, xmm_t0);
+        xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_t1);
+
+        if (len == 0)
+            goto done;
+
+        dst += 32;
+        src += 32;
+    } else if (len + 48 >= 0) {
+        len += 48;
+
+        xmm_t0 = _mm_load_si128((__m128i *)src);
+
+        fold_1(s, &xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3);
+
+        _mm_storeu_si128((__m128i *)dst, xmm_t0);
+
+        xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_t0);
+
+        if (len == 0)
+            goto done;
+
+        dst += 16;
+        src += 16;
+    } else {
+        len += 64;
+        if (len == 0)
+            goto done;
+    }
+
+partial:
+
+#if defined(_MSC_VER)
+    /* VS does not permit the use of _mm_set_epi64x in 32-bit builds */
+    {
+        int32_t parts[4] = {0, 0, 0, 0};
+        memcpy(&parts, src, len);
+        xmm_crc_part = _mm_set_epi32(parts[3], parts[2], parts[1], parts[0]);
+    }
+#else
+    {
+        int64_t parts[2] = {0, 0};
+        memcpy(&parts, src, len);
+        xmm_crc_part = _mm_set_epi64x(parts[1], parts[0]);
+    }
+#endif
+
+    zmemcpy(dst, src, len);  /* TODO: Possibly generate more efficient code. */
+    partial_fold(s, len, &xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3,
+        &xmm_crc_part);
+done:
+    CRC_SAVE(s)
+}
+
+local const unsigned zalign(16) crc_k[] = {
+    0xccaa009e, 0x00000000, /* rk1 */
+    0x751997d0, 0x00000001, /* rk2 */
+    0xccaa009e, 0x00000000, /* rk5 */
+    0x63cd6124, 0x00000001, /* rk6 */
+    0xf7011640, 0x00000001, /* rk7 */
+    0xdb710640, 0x00000001  /* rk8 */
+};
+
+local const unsigned zalign(16) crc_mask[4] = {
+    0xFFFFFFFF, 0xFFFFFFFF, 0x00000000, 0x00000000
+};
+
+local const unsigned zalign(16) crc_mask2[4] = {
+    0x00000000, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF
+};
+
+unsigned ZLIB_INTERNAL crc_fold_512to32(deflate_state *const s)
+{
+    const __m128i xmm_mask  = _mm_load_si128((__m128i *)crc_mask);
+    const __m128i xmm_mask2 = _mm_load_si128((__m128i *)crc_mask2);
+
+    unsigned crc;
+    __m128i x_tmp0, x_tmp1, x_tmp2, crc_fold;
+
+    CRC_LOAD(s)
+
+    /*
+     * k1
+     */
+    crc_fold = _mm_load_si128((__m128i *)crc_k);
+
+    x_tmp0 = _mm_clmulepi64_si128(xmm_crc0, crc_fold, 0x10);
+    xmm_crc0 = _mm_clmulepi64_si128(xmm_crc0, crc_fold, 0x01);
+    xmm_crc1 = _mm_xor_si128(xmm_crc1, x_tmp0);
+    xmm_crc1 = _mm_xor_si128(xmm_crc1, xmm_crc0);
+
+    x_tmp1 = _mm_clmulepi64_si128(xmm_crc1, crc_fold, 0x10);
+    xmm_crc1 = _mm_clmulepi64_si128(xmm_crc1, crc_fold, 0x01);
+    xmm_crc2 = _mm_xor_si128(xmm_crc2, x_tmp1);
+    xmm_crc2 = _mm_xor_si128(xmm_crc2, xmm_crc1);
+
+    x_tmp2 = _mm_clmulepi64_si128(xmm_crc2, crc_fold, 0x10);
+    xmm_crc2 = _mm_clmulepi64_si128(xmm_crc2, crc_fold, 0x01);
+    xmm_crc3 = _mm_xor_si128(xmm_crc3, x_tmp2);
+    xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_crc2);
+
+    /*
+     * k5
+     */
+    crc_fold = _mm_load_si128((__m128i *)crc_k + 1);
+
+    xmm_crc0 = xmm_crc3;
+    xmm_crc3 = _mm_clmulepi64_si128(xmm_crc3, crc_fold, 0);
+    xmm_crc0 = _mm_srli_si128(xmm_crc0, 8);
+    xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_crc0);
+
+    xmm_crc0 = xmm_crc3;
+    xmm_crc3 = _mm_slli_si128(xmm_crc3, 4);
+    xmm_crc3 = _mm_clmulepi64_si128(xmm_crc3, crc_fold, 0x10);
+    xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_crc0);
+    xmm_crc3 = _mm_and_si128(xmm_crc3, xmm_mask2);
+
+    /*
+     * k7
+     */
+    xmm_crc1 = xmm_crc3;
+    xmm_crc2 = xmm_crc3;
+    crc_fold = _mm_load_si128((__m128i *)crc_k + 2);
+
+    xmm_crc3 = _mm_clmulepi64_si128(xmm_crc3, crc_fold, 0);
+    xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_crc2);
+    xmm_crc3 = _mm_and_si128(xmm_crc3, xmm_mask);
+
+    xmm_crc2 = xmm_crc3;
+    xmm_crc3 = _mm_clmulepi64_si128(xmm_crc3, crc_fold, 0x10);
+    xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_crc2);
+    xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_crc1);
+
+    crc = _mm_extract_epi32(xmm_crc3, 2);
+    return ~crc;
+    CRC_SAVE(s)
+}
diff --git a/deflate.c b/deflate.c
index 1ec761448de9..aa0c9c67a6dc 100644
--- a/deflate.c
+++ b/deflate.c
@@ -48,8 +48,9 @@
  */

 /* @(#) $Id$ */
-
+#include <assert.h>
 #include "deflate.h"
+#include "x86.h"

 const char deflate_copyright[] =
    " deflate 1.2.11 Copyright 1995-2017 Jean-loup Gailly and Mark Adler ";
@@ -86,7 +87,7 @@ local block_state deflate_huff   OF((deflate_state *s, int flush));
 local void lm_init        OF((deflate_state *s));
 local void putShortMSB    OF((deflate_state *s, uInt b));
 local void flush_pending  OF((z_streamp strm));
-local unsigned read_buf   OF((z_streamp strm, Bytef *buf, unsigned size));
+unsigned ZLIB_INTERNAL deflate_read_buf OF((z_streamp strm, Bytef *buf, unsigned size));
 #ifdef ASMV
 #  pragma message("Assembler code may have bugs -- use at your own risk")
       void match_init OF((void)); /* asm code initialization */
@@ -100,6 +101,20 @@ local  void check_match OF((deflate_state *s, IPos start, IPos match,
                             int length));
 #endif

+/* From crc32.c */
+extern void ZLIB_INTERNAL crc_reset(deflate_state *const s);
+extern void ZLIB_INTERNAL crc_finalize(deflate_state *const s);
+extern void ZLIB_INTERNAL copy_with_crc(z_streamp strm, Bytef *dst, long size);
+
+#ifdef _MSC_VER
+#define INLINE __inline
+#else
+#define INLINE inline
+#endif
+
+/* Inline optimisation */
+local INLINE Pos insert_string_sse(deflate_state *const s, const Pos str);
+
 /* ===========================================================================
  * Local data
  */
@@ -162,7 +177,6 @@ local const config configuration_table[10] = {
  */
 #define UPDATE_HASH(s,h,c) (h = (((h)<<s->hash_shift) ^ (c)) & s->hash_mask)

-
 /* ===========================================================================
  * Insert string str in the dictionary and set match_head to the previous head
  * of the hash chain (the most recent string with same hash key). Return
@@ -173,17 +187,28 @@ local const config configuration_table[10] = {
  *    characters and the first MIN_MATCH bytes of str are valid (except for
  *    the last MIN_MATCH-1 bytes of the input file).
  */
+local INLINE Pos insert_string_c(deflate_state *const s, const Pos str)
+{
+    Pos ret;
+
+    UPDATE_HASH(s, s->ins_h, s->window[(str) + (MIN_MATCH-1)]);
 #ifdef FASTEST
-#define INSERT_STRING(s, str, match_head) \
-   (UPDATE_HASH(s, s->ins_h, s->window[(str) + (MIN_MATCH-1)]), \
-    match_head = s->head[s->ins_h], \
-    s->head[s->ins_h] = (Pos)(str))
+    ret = s->head[s->ins_h];
 #else
-#define INSERT_STRING(s, str, match_head) \
-   (UPDATE_HASH(s, s->ins_h, s->window[(str) + (MIN_MATCH-1)]), \
-    match_head = s->prev[(str) & s->w_mask] = s->head[s->ins_h], \
-    s->head[s->ins_h] = (Pos)(str))
+    ret = s->prev[str & s->w_mask] = s->head[s->ins_h];
 #endif
+    s->head[s->ins_h] = str;
+
+    return ret;
+}
+
+local INLINE Pos insert_string(deflate_state *const s, const Pos str)
+{
+    if (x86_cpu_enable_simd)
+        return insert_string_sse(s, str);
+    return insert_string_c(s, str);
+}
+

 /* ===========================================================================
  * Initialize the hash table (avoiding 64K overflow for 16 bit systems).
@@ -248,6 +273,7 @@ int ZEXPORT deflateInit2_(strm, level, method, windowBits, memLevel, strategy,
     const char *version;
     int stream_size;
 {
+    unsigned window_padding = 8;
     deflate_state *s;
     int wrap = 1;
     static const char my_version[] = ZLIB_VERSION;
@@ -257,6 +283,8 @@ int ZEXPORT deflateInit2_(strm, level, method, windowBits, memLevel, strategy,
      * output size for (length,distance) codes is <= 24 bits.
      */

+    x86_check_features();
+
     if (version == Z_NULL || version[0] != my_version[0] ||
         stream_size != sizeof(z_stream)) {
         return Z_VERSION_ERROR;
@@ -313,12 +341,19 @@ int ZEXPORT deflateInit2_(strm, level, method, windowBits, memLevel, strategy,
     s->w_size = 1 << s->w_bits;
     s->w_mask = s->w_size - 1;

-    s->hash_bits = (uInt)memLevel + 7;
+    if (x86_cpu_enable_simd) {
+        s->hash_bits = 15;
+    } else {
+        s->hash_bits = memLevel + 7;
+    }
+
     s->hash_size = 1 << s->hash_bits;
     s->hash_mask = s->hash_size - 1;
     s->hash_shift =  ((s->hash_bits+MIN_MATCH-1)/MIN_MATCH);

-    s->window = (Bytef *) ZALLOC(strm, s->w_size, 2*sizeof(Byte));
+    s->window = (Bytef *) ZALLOC(strm,
+                                 s->w_size + window_padding,
+                                 2*sizeof(Byte));
     s->prev   = (Posf *)  ZALLOC(strm, s->w_size, sizeof(Pos));
     s->head   = (Posf *)  ZALLOC(strm, s->hash_size, sizeof(Pos));

@@ -418,11 +453,7 @@ int ZEXPORT deflateSetDictionary (strm, dictionary, dictLength)
         str = s->strstart;
         n = s->lookahead - (MIN_MATCH-1);
         do {
-            UPDATE_HASH(s, s->ins_h, s->window[str + MIN_MATCH-1]);
-#ifndef FASTEST
-            s->prev[str & s->w_mask] = s->head[s->ins_h];
-#endif
-            s->head[s->ins_h] = (Pos)str;
+            insert_string(s, str);
             str++;
         } while (--n);
         s->strstart = str;
@@ -848,7 +879,7 @@ int ZEXPORT deflate (strm, flush)
 #ifdef GZIP
     if (s->status == GZIP_STATE) {
         /* gzip header */
-        strm->adler = crc32(0L, Z_NULL, 0);
+        crc_reset(s);
         put_byte(s, 31);
         put_byte(s, 139);
         put_byte(s, 8);
@@ -1049,6 +1080,7 @@ int ZEXPORT deflate (strm, flush)
     /* Write the trailer */
 #ifdef GZIP
     if (s->wrap == 2) {
+        crc_finalize(s);
         put_byte(s, (Byte)(strm->adler & 0xff));
         put_byte(s, (Byte)((strm->adler >> 8) & 0xff));
         put_byte(s, (Byte)((strm->adler >> 16) & 0xff));
@@ -1161,7 +1193,7 @@ int ZEXPORT deflateCopy (dest, source)
  * allocating a large strm->next_in buffer and copying from it.
  * (See also flush_pending()).
  */
-local unsigned read_buf(strm, buf, size)
+ZLIB_INTERNAL unsigned deflate_read_buf(strm, buf, size)
     z_streamp strm;
     Bytef *buf;
     unsigned size;
@@ -1173,15 +1205,16 @@ local unsigned read_buf(strm, buf, size)

     strm->avail_in  -= len;

-    zmemcpy(buf, strm->next_in, len);
-    if (strm->state->wrap == 1) {
-        strm->adler = adler32(strm->adler, buf, len);
-    }
 #ifdef GZIP
-    else if (strm->state->wrap == 2) {
-        strm->adler = crc32(strm->adler, buf, len);
-    }
+    if (strm->state->wrap == 2)
+        copy_with_crc(strm, buf, len);
+    else
 #endif
+    {
+        zmemcpy(buf, strm->next_in, len);
+        if (strm->state->wrap == 1)
+            strm->adler = adler32(strm->adler, buf, len);
+    }
     strm->next_in  += len;
     strm->total_in += len;

@@ -1479,7 +1512,19 @@ local void check_match(s, start, match, length)
  *    performed for at least two bytes (required for the zip translate_eol
  *    option -- not supported here).
  */
-local void fill_window(s)
+local void fill_window_c(deflate_state *s);
+
+local void fill_window(deflate_state *s)
+{
+    if (x86_cpu_enable_simd) {
+        fill_window_sse(s);
+        return;
+    }
+
+    fill_window_c(s);
+}
+
+local void fill_window_c(s)
     deflate_state *s;
 {
     unsigned n;
@@ -1847,7 +1892,7 @@ local block_state deflate_fast(s, flush)
          */
         hash_head = NIL;
         if (s->lookahead >= MIN_MATCH) {
-            INSERT_STRING(s, s->strstart, hash_head);
+            hash_head = insert_string(s, s->strstart);
         }

         /* Find the longest match, discarding those <= prev_length.
@@ -1878,7 +1923,7 @@ local block_state deflate_fast(s, flush)
                 s->match_length--; /* string at strstart already in table */
                 do {
                     s->strstart++;
-                    INSERT_STRING(s, s->strstart, hash_head);
+                    hash_head = insert_string(s, s->strstart);
                     /* strstart never exceeds WSIZE-MAX_MATCH, so there are
                      * always MIN_MATCH bytes ahead.
                      */
@@ -1950,7 +1995,7 @@ local block_state deflate_slow(s, flush)
          */
         hash_head = NIL;
         if (s->lookahead >= MIN_MATCH) {
-            INSERT_STRING(s, s->strstart, hash_head);
+            hash_head = insert_string(s, s->strstart);
         }

         /* Find the longest match, discarding those <= prev_length.
@@ -2001,7 +2046,7 @@ local block_state deflate_slow(s, flush)
             s->prev_length -= 2;
             do {
                 if (++s->strstart <= max_insert) {
-                    INSERT_STRING(s, s->strstart, hash_head);
+                    hash_head = insert_string(s, s->strstart);
                 }
             } while (--s->prev_length != 0);
             s->match_available = 0;
@@ -2161,3 +2206,37 @@ local block_state deflate_huff(s, flush)
         FLUSH_BLOCK(s, 0);
     return block_done;
 }
+
+/* Safe to inline this as GCC/clang will use inline asm and Visual Studio will
+ * use intrinsic without extra params
+ */
+local INLINE Pos insert_string_sse(deflate_state *const s, const Pos str)
+{
+    Pos ret;
+    unsigned *ip, val, h = 0;
+
+    ip = (unsigned *)&s->window[str];
+    val = *ip;
+
+    if (s->level >= 6)
+        val &= 0xFFFFFF;
+
+/* Windows clang should use inline asm */
+#if defined(_MSC_VER) && !defined(__clang__)
+    h = _mm_crc32_u32(h, val);
+#elif defined(__i386__) || defined(__amd64__)
+    __asm__ __volatile__ (
+        "crc32 %1,%0\n\t"
+    : "+r" (h)
+    : "r" (val)
+    );
+#else
+    /* This should never happen */
+    assert(0);
+#endif
+
+    ret = s->head[h & s->hash_mask];
+    s->head[h & s->hash_mask] = str;
+    s->prev[str & s->w_mask] = ret;
+    return ret;
+}
diff --git a/deflate.h b/deflate.h
index 23ecdd312bc0..ab56df7663b6 100644
--- a/deflate.h
+++ b/deflate.h
@@ -109,7 +109,7 @@ typedef struct internal_state {
     ulg   gzindex;       /* where in extra, name, or comment */
     Byte  method;        /* can only be DEFLATED */
     int   last_flush;    /* value of flush param for previous deflate call */
-
+    unsigned zalign(16) crc0[4 * 5];
                 /* used by deflate.c: */

     uInt  w_size;        /* LZ77 window size (32K by default) */
@@ -346,4 +346,14 @@ void ZLIB_INTERNAL _tr_stored_block OF((deflate_state *s, charf *buf,
               flush = _tr_tally(s, distance, length)
 #endif

+/* Functions that are SIMD optimised on x86 */
+void ZLIB_INTERNAL crc_fold_init(deflate_state* const s);
+void ZLIB_INTERNAL crc_fold_copy(deflate_state* const s,
+                                 unsigned char* dst,
+                                 const unsigned char* src,
+                                 long len);
+unsigned ZLIB_INTERNAL crc_fold_512to32(deflate_state* const s);
+
+void ZLIB_INTERNAL fill_window_sse(deflate_state* s);
+
 #endif /* DEFLATE_H */
diff --git a/fill_window_sse.c b/fill_window_sse.c
new file mode 100644
index 000000000000..949ccce1ba9c
--- /dev/null
+++ b/fill_window_sse.c
@@ -0,0 +1,177 @@
+/*
+ * Fill Window with SSE2-optimized hash shifting
+ *
+ * Copyright (C) 2013 Intel Corporation
+ * Authors:
+ *  Arjan van de Ven    <[email protected]>
+ *  Jim Kukunas         <[email protected]>
+ *
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+#include <immintrin.h>
+#include "deflate.h"
+
+#define UPDATE_HASH(s,h,i) \
+    {\
+        if (s->level < 6) { \
+            h = (3483 * (s->window[i]) +\
+                 23081* (s->window[i+1]) +\
+                 6954 * (s->window[i+2]) +\
+                 20947* (s->window[i+3])) & s->hash_mask;\
+        } else {\
+            h = (25881* (s->window[i]) +\
+                 24674* (s->window[i+1]) +\
+                 25811* (s->window[i+2])) & s->hash_mask;\
+        }\
+    }\
+
+extern int deflate_read_buf OF((z_streamp strm, Bytef *buf, unsigned size));
+
+void fill_window_sse(deflate_state *s)
+{
+    const __m128i xmm_wsize = _mm_set1_epi16(s->w_size);
+
+    register unsigned n;
+    register Posf *p;
+    unsigned more;    /* Amount of free space at the end of the window. */
+    uInt wsize = s->w_size;
+
+    Assert(s->lookahead < MIN_LOOKAHEAD, "already enough lookahead");
+
+    do {
+        more = (unsigned)(s->window_size -(ulg)s->lookahead -(ulg)s->strstart);
+
+        /* Deal with !@#$% 64K limit: */
+        if (sizeof(int) <= 2) {
+            if (more == 0 && s->strstart == 0 && s->lookahead == 0) {
+                more = wsize;
+
+            } else if (more == (unsigned)(-1)) {
+                /* Very unlikely, but possible on 16 bit machine if
+                 * strstart == 0 && lookahead == 1 (input done a byte at time)
+                 */
+                more--;
+            }
+        }
+
+        /* If the window is almost full and there is insufficient lookahead,
+         * move the upper half to the lower one to make room in the upper half.
+         */
+        if (s->strstart >= wsize+MAX_DIST(s)) {
+
+            zmemcpy(s->window, s->window+wsize, (unsigned)wsize);
+            s->match_start -= wsize;
+            s->strstart    -= wsize; /* we now have strstart >= MAX_DIST */
+            s->block_start -= (long) wsize;
+
+            /* Slide the hash table (could be avoided with 32 bit values
+               at the expense of memory usage). We slide even when level == 0
+               to keep the hash table consistent if we switch back to level > 0
+               later. (Using level 0 permanently is not an optimal usage of
+               zlib, so we don't care about this pathological case.)
+             */
+            n = s->hash_size;
+            p = &s->head[n];
+            p -= 8;
+            do {
+                __m128i value, result;
+
+                value = _mm_loadu_si128((__m128i *)p);
+                result = _mm_subs_epu16(value, xmm_wsize);
+                _mm_storeu_si128((__m128i *)p, result);
+
+                p -= 8;
+                n -= 8;
+            } while (n > 0);
+
+            n = wsize;
+#ifndef FASTEST
+            p = &s->prev[n];
+            p -= 8;
+            do {
+                __m128i value, result;
+
+                value = _mm_loadu_si128((__m128i *)p);
+                result = _mm_subs_epu16(value, xmm_wsize);
+                _mm_storeu_si128((__m128i *)p, result);
+
+                p -= 8;
+                n -= 8;
+            } while (n > 0);
+#endif
+            more += wsize;
+        }
+        if (s->strm->avail_in == 0) break;
+
+        /* If there was no sliding:
+         *    strstart <= WSIZE+MAX_DIST-1 && lookahead <= MIN_LOOKAHEAD - 1 &&
+         *    more == window_size - lookahead - strstart
+         * => more >= window_size - (MIN_LOOKAHEAD-1 + WSIZE + MAX_DIST-1)
+         * => more >= window_size - 2*WSIZE + 2
+         * In the BIG_MEM or MMAP case (not yet supported),
+         *   window_size == input_size + MIN_LOOKAHEAD  &&
+         *   strstart + s->lookahead <= input_size => more >= MIN_LOOKAHEAD.
+         * Otherwise, window_size == 2*WSIZE so more >= 2.
+         * If there was sliding, more >= WSIZE. So in all cases, more >= 2.
+         */
+        Assert(more >= 2, "more < 2");
+
+        n = deflate_read_buf(s->strm,
+                             s->window + s->strstart + s->lookahead,
+                             more);
+        s->lookahead += n;
+
+        /* Initialize the hash value now that we have some input: */
+        if (s->lookahead >= MIN_MATCH) {
+            uInt str = s->strstart;
+            s->ins_h = s->window[str];
+            if (str >= 1)
+                UPDATE_HASH(s, s->ins_h, str + 1 - (MIN_MATCH-1));
+#if MIN_MATCH != 3
+            Call UPDATE_HASH() MIN_MATCH-3 more times
+#endif
+        }
+        /* If the whole input has less than MIN_MATCH bytes, ins_h is garbage,
+         * but this is not important since only literal bytes will be emitted.
+         */
+
+    } while (s->lookahead < MIN_LOOKAHEAD && s->strm->avail_in != 0);
+
+    /* If the WIN_INIT bytes after the end of the current data have never been
+     * written, then zero those bytes in order to avoid memory check reports of
+     * the use of uninitialized (or uninitialised as Julian writes) bytes by
+     * the longest match routines.  Update the high water mark for the next
+     * time through here.  WIN_INIT is set to MAX_MATCH since the longest match
+     * routines allow scanning to strstart + MAX_MATCH, ignoring lookahead.
+     */
+    if (s->high_water < s->window_size) {
+        ulg curr = s->strstart + (ulg)(s->lookahead);
+        ulg init;
+
+        if (s->high_water < curr) {
+            /* Previous high water mark below current data -- zero WIN_INIT
+             * bytes or up to end of window, whichever is less.
+             */
+            init = s->window_size - curr;
+            if (init > WIN_INIT)
+                init = WIN_INIT;
+            zmemzero(s->window + curr, (unsigned)init);
+            s->high_water = curr + init;
+        }
+        else if (s->high_water < (ulg)curr + WIN_INIT) {
+            /* High water mark at or above current data, but below current data
+             * plus WIN_INIT -- zero out to current data plus WIN_INIT, or up
+             * to end of window, whichever is less.
+             */
+            init = (ulg)curr + WIN_INIT - s->high_water;
+            if (init > s->window_size - s->high_water)
+                init = s->window_size - s->high_water;
+            zmemzero(s->window + s->high_water, (unsigned)init);
+            s->high_water += init;
+        }
+    }
+
+    Assert((ulg)s->strstart <= s->window_size - MIN_LOOKAHEAD,
+           "not enough room for search");
+}
diff --git a/simd_stub.c b/simd_stub.c
new file mode 100644
index 000000000000..c6d46051498f
--- /dev/null
+++ b/simd_stub.c
@@ -0,0 +1,35 @@
+/* simd_stub.c -- stub implementations
+* Copyright (C) 2014 Intel Corporation
+* For conditions of distribution and use, see copyright notice in zlib.h
+*/
+#include <assert.h>
+
+#include "deflate.h"
+#include "x86.h"
+
+int ZLIB_INTERNAL x86_cpu_enable_simd = 0;
+
+void ZLIB_INTERNAL crc_fold_init(deflate_state *const s) {
+    assert(0);
+}
+
+void ZLIB_INTERNAL crc_fold_copy(deflate_state *const s,
+                                 unsigned char *dst,
+                                 const unsigned char *src,
+                                 long len) {
+    assert(0);
+}
+
+unsigned ZLIB_INTERNAL crc_fold_512to32(deflate_state *const s) {
+    assert(0);
+    return 0;
+}
+
+void ZLIB_INTERNAL fill_window_sse(deflate_state *s)
+{
+    assert(0);
+}
+
+void x86_check_features(void)
+{
+}
diff --git a/x86.c b/x86.c
new file mode 100644
index 000000000000..e56fe8b85a39
--- /dev/null
+++ b/x86.c
@@ -0,0 +1,92 @@
+/*
+ * x86 feature check
+ *
+ * Copyright (C) 2013 Intel Corporation. All rights reserved.
+ * Author:
+ *  Jim Kukunas
+ *
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+#include "x86.h"
+#include "zutil.h"
+
+int ZLIB_INTERNAL x86_cpu_enable_simd = 0;
+
+#ifndef _MSC_VER
+#include <pthread.h>
+
+pthread_once_t cpu_check_inited_once = PTHREAD_ONCE_INIT;
+static void _x86_check_features(void);
+
+void x86_check_features(void)
+{
+  pthread_once(&cpu_check_inited_once, _x86_check_features);
+}
+
+static void _x86_check_features(void)
+{
+    int x86_cpu_has_sse2;
+    int x86_cpu_has_sse42;
+    int x86_cpu_has_pclmulqdq;
+    unsigned eax, ebx, ecx, edx;
+
+    eax = 1;
+#ifdef __i386__
+    __asm__ __volatile__ (
+        "xchg %%ebx, %1\n\t"
+        "cpuid\n\t"
+        "xchg %1, %%ebx\n\t"
+    : "+a" (eax), "=S" (ebx), "=c" (ecx), "=d" (edx)
+    );
+#else
+    __asm__ __volatile__ (
+        "cpuid\n\t"
+    : "+a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx)
+    );
+#endif  /* (__i386__) */
+
+    x86_cpu_has_sse2 = edx & 0x4000000;
+    x86_cpu_has_sse42 = ecx & 0x100000;
+    x86_cpu_has_pclmulqdq = ecx & 0x2;
+
+    x86_cpu_enable_simd = x86_cpu_has_sse2 &&
+                          x86_cpu_has_sse42 &&
+                          x86_cpu_has_pclmulqdq;
+}
+#else
+#include <intrin.h>
+#include <windows.h>
+
+static BOOL CALLBACK _x86_check_features(PINIT_ONCE once,
+                                         PVOID param,
+                                         PVOID *context);
+static INIT_ONCE cpu_check_inited_once = INIT_ONCE_STATIC_INIT;
+
+void x86_check_features(void)
+{
+    InitOnceExecuteOnce(&cpu_check_inited_once, _x86_check_features,
+                        NULL, NULL);
+}
+
+static BOOL CALLBACK _x86_check_features(PINIT_ONCE once,
+                                         PVOID param,
+                                         PVOID *context)
+{
+    int x86_cpu_has_sse2;
+    int x86_cpu_has_sse42;
+    int x86_cpu_has_pclmulqdq;
+    int regs[4];
+
+    __cpuid(regs, 1);
+
+    x86_cpu_has_sse2 = regs[3] & 0x4000000;
+    x86_cpu_has_sse42= regs[2] & 0x100000;
+    x86_cpu_has_pclmulqdq = regs[2] & 0x2;
+
+    x86_cpu_enable_simd = x86_cpu_has_sse2 &&
+                          x86_cpu_has_sse42 &&
+                          x86_cpu_has_pclmulqdq;
+    return TRUE;
+}
+#endif  /* _MSC_VER */
diff --git a/x86.h b/x86.h
new file mode 100644
index 000000000000..ebcf10ab09d2
--- /dev/null
+++ b/x86.h
@@ -0,0 +1,15 @@
+/* x86.h -- check for x86 CPU features
+* Copyright (C) 2013 Intel Corporation Jim Kukunas
+* For conditions of distribution and use, see copyright notice in zlib.h
+*/
+
+#ifndef X86_H
+#define X86_H
+
+#include "zlib.h"
+
+extern int x86_cpu_enable_simd;
+
+void x86_check_features(void);
+
+#endif  /* X86_H */
diff --git a/zutil.h b/zutil.h
index 80375b8b6109..4425bcf75eb3 100644
--- a/zutil.h
+++ b/zutil.h
@@ -283,4 +283,10 @@ extern z_const char * const z_errmsg[10]; /* indexed by 2-zlib_error */
 #define ZSWAP32(q) ((((q) >> 24) & 0xff) + (((q) >> 8) & 0xff00) + \
                     (((q) & 0xff00) << 8) + (((q) & 0xff) << 24))

+#ifdef _MSC_VER
+#define zalign(x) __declspec(align(x))
+#else
+#define zalign(x) __attribute__((aligned((x))))
+#endif
+
 #endif /* ZUTIL_H */