lib/libz/0003-Port-Intel-optimizations-adler32-chunkcopy-to-cloudf.patch
$ cat 0003-Port-Intel-optimizations-adler32-chunkcopy-to-cloudf.patch
From 7b243be132503e86d9be0e18454bb05ca033118f Mon Sep 17 00:00:00 2001
From: Janakarajan Natarajan <68447808+janaknat@users.noreply.github.com>
Date: Mon, 28 Sep 2020 13:24:49 -0500
Subject: [PATCH 3/4] Port Intel optimizations (adler32, chunkcopy) to
cloudflare (#23)
* Add SIMD SSSE3 implementation of the adler32 checksum
Based on the adler32-simd patch from Noel Gordon for the chromium fork of zlib.
17bbb3d73c84 ("zlib adler_simd.c")
Signed-off-by: Janakarajan Natarajan <janakan@amazon.com>
* Port inflate chunk SIMD SSE2 improvements for cloudflare
Based on 2 patches from zlib chromium fork:
* Adenilson Cavalcanti (adenilson.cavalcanti@arm.com)
3060dcb - "zlib: inflate using wider loads and stores"
* Noel Gordon (noel@chromium.org)
64ffef0 - "Improve zlib inflate speed by using SSE2 chunk copy
The improvement in inflate performance is around 15-35%, based
on the workload, when checked with a modified zpipe.c and the
Silesia corpus.
Signed-off-by: Janakarajan Natarajan <janakan@amazon.com>
---
Makefile.in | 12 ++
adler32.c | 9 ++
adler32_simd.c | 387 ++++++++++++++++++++++++++++++++++++++++++++++++
adler32_simd.h | 37 +++++
chunkcopy.h | 111 ++++++++++----
configure | 26 ++++
inffast_chunk.c | 24 +--
inffast_chunk.h | 2 +-
inflate.c | 10 +-
9 files changed, 570 insertions(+), 48 deletions(-)
create mode 100644 adler32_simd.c
create mode 100644 adler32_simd.h
diff --git a/Makefile.in b/Makefile.in
index 3647d90..fde5405 100644
--- a/Makefile.in
+++ b/Makefile.in
@@ -41,6 +41,18 @@ ifneq ($(findstring -DINFLATE_CHUNK_SIMD_NEON,$(CFLAGS)),)
STATIC_OBJS += inffast_chunk.o
endif
+ifneq ($(findstring -DINFLATE_CHUNK_SIMD_SSE2,$(CFLAGS)),)
+STATIC_OBJS += inffast_chunk.o
+endif
+
+ifneq ($(findstring -DADLER32_SIMD_NEON,$(CFLAGS)),)
+STATIC_OBJS += adler32_simd.o
+endif
+
+ifneq ($(findstring -DADLER32_SIMD_SSSE3,$(CFLAGS)),)
+STATIC_OBJS += adler32_simd.o
+endif
+
# TODO: What extension to use here?
SHARED_OBJS=$(STATIC_OBJS:.o=.lo)
diff --git a/adler32.c b/adler32.c
index 127a9a5..05da2ae 100644
--- a/adler32.c
+++ b/adler32.c
@@ -10,6 +10,10 @@
#include "zutil.h"
+#if defined(ADLER32_SIMD_NEON) || defined(ADLER32_SIMD_SSSE3)
+#include "adler32_simd.h"
+#endif
+
#define BASE 65521 /* largest prime smaller than 65536 */
#define BASE_X2 131042 /* twice BASE */
#define NMAX 5552
@@ -20,6 +24,11 @@ unsigned long ZEXPORT adler32(unsigned long adler,
const unsigned char *buf,
unsigned int len)
{
+#if defined(ADLER32_SIMD_NEON) || defined(ADLER32_SIMD_SSSE3)
+ if (buf != NULL && len >= 64)
+ return adler32_simd_((uint32_t)adler, buf, len);
+#endif
+
/* initial Adler-32 value */
if (buf == NULL)
return UINT32_C(1) << 0 | UINT32_C(0) << 16;
diff --git a/adler32_simd.c b/adler32_simd.c
new file mode 100644
index 0000000..5f8c06d
--- /dev/null
+++ b/adler32_simd.c
@@ -0,0 +1,387 @@
+/* adler32_simd.c
+ *
+ * (C) 1995-2013 Jean-loup Gailly and Mark Adler
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty. In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ * claim that you wrote the original software. If you use this software
+ * in a product, an acknowledgment in the product documentation would be
+ * appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ * misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ *
+ * Jean-loup Gailly Mark Adler
+ * jloup@gzip.org madler@alumni.caltech.edu
+ *
+ * Copyright 2017 The Chromium Authors. All rights reserved.
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the Chromium source repository LICENSE file.
+ *
+ * Per http://en.wikipedia.org/wiki/Adler-32 the adler32 A value (aka s1) is
+ * the sum of N input data bytes D1 ... DN,
+ *
+ * A = A0 + D1 + D2 + ... + DN
+ *
+ * where A0 is the initial value.
+ *
+ * SSE2 _mm_sad_epu8() can be used for byte sums (see http://bit.ly/2wpUOeD,
+ * for example) and accumulating the byte sums can use SSE shuffle-adds (see
+ * the "Integer" section of http://bit.ly/2erPT8t for details). Arm NEON has
+ * similar instructions.
+ *
+ * The adler32 B value (aka s2) sums the A values from each step:
+ *
+ * B0 + (A0 + D1) + (A0 + D1 + D2) + ... + (A0 + D1 + D2 + ... + DN) or
+ *
+ * B0 + N.A0 + N.D1 + (N-1).D2 + (N-2).D3 + ... + (N-(N-1)).DN
+ *
+ * B0 being the initial value. For 32 bytes (ideal for garden-variety SIMD):
+ *
+ * B = B0 + 32.A0 + [D1 D2 D3 ... D32] x [32 31 30 ... 1].
+ *
+ * Adjacent blocks of 32 input bytes can be iterated with the expressions to
+ * compute the adler32 s1 s2 of M >> 32 input bytes [1].
+ *
+ * As M grows, the s1 s2 sums grow. If left unchecked, they would eventually
+ * overflow the precision of their integer representation (bad). However, s1
+ * and s2 also need to be computed modulo the adler BASE value (reduced). If
+ * at most NMAX bytes are processed before a reduce, s1 s2 _cannot_ overflow
+ * a uint32_t type (the NMAX constraint) [2].
+ *
+ * [1] the iterative equations for s2 contain constant factors; these can be
+ * hoisted from the n-blocks do loop of the SIMD code.
+ *
+ * [2] zlib adler32_z() uses this fact to implement NMAX-block-based updates
+ * of the adler s1 s2 of uint32_t type (see adler32.c).
+ */
+
+#include "adler32_simd.h"
+
+/* Definitions from adler32.c: largest prime smaller than 65536 */
+#define BASE 65521U
+/* NMAX is the largest n such that 255n(n+1)/2 + (n+1)(BASE-1) <= 2^32-1 */
+#define NMAX 5552
+
+#if defined(ADLER32_SIMD_SSSE3)
+
+#include <tmmintrin.h>
+
+uint32_t ZLIB_INTERNAL adler32_simd_( /* SSSE3 */
+ uint32_t adler,
+ const unsigned char *buf,
+ unsigned long len)
+{
+ /*
+ * Split Adler-32 into component sums.
+ */
+ uint32_t s1 = adler & 0xffff;
+ uint32_t s2 = adler >> 16;
+
+ /*
+ * Process the data in blocks.
+ */
+ const unsigned BLOCK_SIZE = 1 << 5;
+
+ unsigned long blocks = len / BLOCK_SIZE;
+ len -= blocks * BLOCK_SIZE;
+
+ while (blocks)
+ {
+ unsigned n = NMAX / BLOCK_SIZE; /* The NMAX constraint. */
+ if (n > blocks)
+ n = (unsigned) blocks;
+ blocks -= n;
+
+ const __m128i tap1 =
+ _mm_setr_epi8(32,31,30,29,28,27,26,25,24,23,22,21,20,19,18,17);
+ const __m128i tap2 =
+ _mm_setr_epi8(16,15,14,13,12,11,10, 9, 8, 7, 6, 5, 4, 3, 2, 1);
+ const __m128i zero =
+ _mm_setr_epi8( 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
+ const __m128i ones =
+ _mm_set_epi16( 1, 1, 1, 1, 1, 1, 1, 1);
+
+ /*
+ * Process n blocks of data. At most NMAX data bytes can be
+ * processed before s2 must be reduced modulo BASE.
+ */
+ __m128i v_ps = _mm_set_epi32(0, 0, 0, s1 * n);
+ __m128i v_s2 = _mm_set_epi32(0, 0, 0, s2);
+ __m128i v_s1 = _mm_set_epi32(0, 0, 0, 0);
+
+ do {
+ /*
+ * Load 32 input bytes.
+ */
+ const __m128i bytes1 = _mm_loadu_si128((__m128i*)(buf));
+ const __m128i bytes2 = _mm_loadu_si128((__m128i*)(buf + 16));
+
+ /*
+ * Add previous block byte sum to v_ps.
+ */
+ v_ps = _mm_add_epi32(v_ps, v_s1);
+
+ /*
+ * Horizontally add the bytes for s1, multiply-adds the
+ * bytes by [ 32, 31, 30, ... ] for s2.
+ */
+ v_s1 = _mm_add_epi32(v_s1, _mm_sad_epu8(bytes1, zero));
+ const __m128i mad1 = _mm_maddubs_epi16(bytes1, tap1);
+ v_s2 = _mm_add_epi32(v_s2, _mm_madd_epi16(mad1, ones));
+
+ v_s1 = _mm_add_epi32(v_s1, _mm_sad_epu8(bytes2, zero));
+ const __m128i mad2 = _mm_maddubs_epi16(bytes2, tap2);
+ v_s2 = _mm_add_epi32(v_s2, _mm_madd_epi16(mad2, ones));
+
+ buf += BLOCK_SIZE;
+
+ } while (--n);
+
+ v_s2 = _mm_add_epi32(v_s2, _mm_slli_epi32(v_ps, 5));
+
+ /*
+ * Sum epi32 ints v_s1(s2) and accumulate in s1(s2).
+ */
+
+#define S23O1 _MM_SHUFFLE(2,3,0,1) /* A B C D -> B A D C */
+#define S1O32 _MM_SHUFFLE(1,0,3,2) /* A B C D -> C D A B */
+
+ v_s1 = _mm_add_epi32(v_s1, _mm_shuffle_epi32(v_s1, S23O1));
+ v_s1 = _mm_add_epi32(v_s1, _mm_shuffle_epi32(v_s1, S1O32));
+
+ s1 += _mm_cvtsi128_si32(v_s1);
+
+ v_s2 = _mm_add_epi32(v_s2, _mm_shuffle_epi32(v_s2, S23O1));
+ v_s2 = _mm_add_epi32(v_s2, _mm_shuffle_epi32(v_s2, S1O32));
+
+ s2 = _mm_cvtsi128_si32(v_s2);
+
+#undef S23O1
+#undef S1O32
+
+ /*
+ * Reduce.
+ */
+ s1 %= BASE;
+ s2 %= BASE;
+ }
+
+ /*
+ * Handle leftover data.
+ */
+ if (len) {
+ if (len >= 16) {
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+
+ len -= 16;
+ }
+
+ while (len--) {
+ s2 += (s1 += *buf++);
+ }
+
+ if (s1 >= BASE)
+ s1 -= BASE;
+ s2 %= BASE;
+ }
+
+ /*
+ * Return the recombined sums.
+ */
+ return s1 | (s2 << 16);
+}
+
+#elif defined(ADLER32_SIMD_NEON)
+
+#include <arm_neon.h>
+
+uint32_t ZLIB_INTERNAL adler32_simd_( /* NEON */
+ uint32_t adler,
+ const unsigned char *buf,
+ unsigned long len)
+{
+ /*
+ * Split Adler-32 into component sums.
+ */
+ uint32_t s1 = adler & 0xffff;
+ uint32_t s2 = adler >> 16;
+
+ /*
+ * Serially compute s1 & s2, until the data is 16-byte aligned.
+ */
+ if ((uintptr_t)buf & 15) {
+ while ((uintptr_t)buf & 15) {
+ s2 += (s1 += *buf++);
+ --len;
+ }
+
+ if (s1 >= BASE)
+ s1 -= BASE;
+ s2 %= BASE;
+ }
+
+ /*
+ * Process the data in blocks.
+ */
+ const unsigned BLOCK_SIZE = 1 << 5;
+
+ unsigned long blocks = len / BLOCK_SIZE;
+ len -= blocks * BLOCK_SIZE;
+
+ while (blocks)
+ {
+ unsigned n = NMAX / BLOCK_SIZE; /* The NMAX constraint. */
+ if (n > blocks)
+ n = blocks;
+ blocks -= n;
+
+ /*
+ * Process n blocks of data. At most NMAX data bytes can be
+ * processed before s2 must be reduced modulo BASE.
+ */
+ uint32x4_t v_s2 = (uint32x4_t) { 0, 0, 0, s1 * n };
+ uint32x4_t v_s1 = (uint32x4_t) { 0, 0, 0, 0 };
+
+ uint16x8_t v_column_sum_1 = vdupq_n_u16(0);
+ uint16x8_t v_column_sum_2 = vdupq_n_u16(0);
+ uint16x8_t v_column_sum_3 = vdupq_n_u16(0);
+ uint16x8_t v_column_sum_4 = vdupq_n_u16(0);
+
+ do {
+ /*
+ * Load 32 input bytes.
+ */
+ const uint8x16_t bytes1 = vld1q_u8((uint8_t*)(buf));
+ const uint8x16_t bytes2 = vld1q_u8((uint8_t*)(buf + 16));
+
+ /*
+ * Add previous block byte sum to v_s2.
+ */
+ v_s2 = vaddq_u32(v_s2, v_s1);
+
+ /*
+ * Horizontally add the bytes for s1.
+ */
+ v_s1 = vpadalq_u16(v_s1, vpadalq_u8(vpaddlq_u8(bytes1), bytes2));
+
+ /*
+ * Vertically add the bytes for s2.
+ */
+ v_column_sum_1 = vaddw_u8(v_column_sum_1, vget_low_u8 (bytes1));
+ v_column_sum_2 = vaddw_u8(v_column_sum_2, vget_high_u8(bytes1));
+ v_column_sum_3 = vaddw_u8(v_column_sum_3, vget_low_u8 (bytes2));
+ v_column_sum_4 = vaddw_u8(v_column_sum_4, vget_high_u8(bytes2));
+
+ buf += BLOCK_SIZE;
+
+ } while (--n);
+
+ v_s2 = vshlq_n_u32(v_s2, 5);
+
+ /*
+ * Multiply-add bytes by [ 32, 31, 30, ... ] for s2.
+ */
+ v_s2 = vmlal_u16(v_s2, vget_low_u16 (v_column_sum_1),
+ (uint16x4_t) { 32, 31, 30, 29 });
+ v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_1),
+ (uint16x4_t) { 28, 27, 26, 25 });
+ v_s2 = vmlal_u16(v_s2, vget_low_u16 (v_column_sum_2),
+ (uint16x4_t) { 24, 23, 22, 21 });
+ v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_2),
+ (uint16x4_t) { 20, 19, 18, 17 });
+ v_s2 = vmlal_u16(v_s2, vget_low_u16 (v_column_sum_3),
+ (uint16x4_t) { 16, 15, 14, 13 });
+ v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_3),
+ (uint16x4_t) { 12, 11, 10, 9 });
+ v_s2 = vmlal_u16(v_s2, vget_low_u16 (v_column_sum_4),
+ (uint16x4_t) { 8, 7, 6, 5 });
+ v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_4),
+ (uint16x4_t) { 4, 3, 2, 1 });
+
+ /*
+ * Sum epi32 ints v_s1(s2) and accumulate in s1(s2).
+ */
+ uint32x2_t sum1 = vpadd_u32(vget_low_u32(v_s1), vget_high_u32(v_s1));
+ uint32x2_t sum2 = vpadd_u32(vget_low_u32(v_s2), vget_high_u32(v_s2));
+ uint32x2_t s1s2 = vpadd_u32(sum1, sum2);
+
+ s1 += vget_lane_u32(s1s2, 0);
+ s2 += vget_lane_u32(s1s2, 1);
+
+ /*
+ * Reduce.
+ */
+ s1 %= BASE;
+ s2 %= BASE;
+ }
+
+ /*
+ * Handle leftover data.
+ */
+ if (len) {
+ if (len >= 16) {
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+ s2 += (s1 += *buf++);
+
+ len -= 16;
+ }
+
+ while (len--) {
+ s2 += (s1 += *buf++);
+ }
+
+ if (s1 >= BASE)
+ s1 -= BASE;
+ s2 %= BASE;
+ }
+
+ /*
+ * Return the recombined sums.
+ */
+ return s1 | (s2 << 16);
+}
+
+#endif /* ADLER32_SIMD_SSSE3 */
diff --git a/adler32_simd.h b/adler32_simd.h
new file mode 100644
index 0000000..285e392
--- /dev/null
+++ b/adler32_simd.h
@@ -0,0 +1,37 @@
+/* adler32_simd.h
+ *
+ * (C) 1995-2013 Jean-loup Gailly and Mark Adler
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty. In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ * claim that you wrote the original software. If you use this software
+ * in a product, an acknowledgment in the product documentation would be
+ * appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ * misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ *
+ * Jean-loup Gailly Mark Adler
+ * jloup@gzip.org madler@alumni.caltech.edu
+ *
+ * Copyright 2017 The Chromium Authors. All rights reserved.
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the Chromium source repository LICENSE file.
+ */
+
+#include <stdint.h>
+
+#include "zconf.h"
+#include "zutil.h"
+
+uint32_t ZLIB_INTERNAL adler32_simd_(
+ uint32_t adler,
+ const unsigned char *buf,
+ unsigned long len);
diff --git a/chunkcopy.h b/chunkcopy.h
index 217da52..cf2b708 100644
--- a/chunkcopy.h
+++ b/chunkcopy.h
@@ -50,6 +50,11 @@
#if defined(INFLATE_CHUNK_SIMD_NEON)
#include <arm_neon.h>
typedef uint8x16_t z_vec128i_t;
+#elif defined(INFLATE_CHUNK_SIMD_SSE2)
+#include <emmintrin.h>
+typedef __m128i z_vec128i_t;
+#else
+#error chunkcopy.h inflate chunk SIMD is not defined for your build target
#endif
/*
@@ -66,7 +71,7 @@ Z_STATIC_ASSERT(vector_128_bits_wide,
* instruction appropriate for the z_vec128i_t type.
*/
static inline z_vec128i_t loadchunk(
- const unsigned char FAR* s) {
+ const unsigned char * s) {
z_vec128i_t v;
Z_BUILTIN_MEMCPY(&v, s, sizeof(v));
return v;
@@ -77,7 +82,7 @@ static inline z_vec128i_t loadchunk(
* instruction appropriate for the z_vec128i_t type.
*/
static inline void storechunk(
- unsigned char FAR* d,
+ unsigned char * d,
const z_vec128i_t v) {
Z_BUILTIN_MEMCPY(d, &v, sizeof(v));
}
@@ -96,9 +101,9 @@ static inline void storechunk(
* without iteration, which will hopefully make the branch prediction more
* reliable.
*/
-static inline unsigned char FAR* chunkcopy_core(
- unsigned char FAR* out,
- const unsigned char FAR* from,
+static inline unsigned char * chunkcopy_core(
+ unsigned char * out,
+ const unsigned char * from,
unsigned len) {
const int bump = (--len % CHUNKCOPY_CHUNK_SIZE) + 1;
storechunk(out, loadchunk(from));
@@ -121,14 +126,14 @@ static inline unsigned char FAR* chunkcopy_core(
* output buffer is beyond the end of the current copy, and this can still be
* exploited.
*/
-static inline unsigned char FAR* chunkcopy_core_safe(
- unsigned char FAR* out,
- const unsigned char FAR* from,
+static inline unsigned char * chunkcopy_core_safe(
+ unsigned char * out,
+ const unsigned char * from,
unsigned len,
- unsigned char FAR* limit) {
+ unsigned char * limit) {
Assert(out + len <= limit, "chunk copy exceeds safety limit");
if ((limit - out) < (ptrdiff_t)CHUNKCOPY_CHUNK_SIZE) {
- const unsigned char FAR* Z_RESTRICT rfrom = from;
+ const unsigned char * Z_RESTRICT rfrom = from;
if (len & 8) {
Z_BUILTIN_MEMCPY(out, rfrom, 8);
out += 8;
@@ -162,11 +167,11 @@ static inline unsigned char FAR* chunkcopy_core_safe(
* least 258 bytes of output space available (258 being the maximum length
* output from a single token; see inffast.c).
*/
-static inline unsigned char FAR* chunkunroll_relaxed(
- unsigned char FAR* out,
- unsigned FAR* dist,
- unsigned FAR* len) {
- const unsigned char FAR* from = out - *dist;
+static inline unsigned char * chunkunroll_relaxed(
+ unsigned char * out,
+ unsigned * dist,
+ unsigned * len) {
+ const unsigned char * from = out - *dist;
while (*dist < *len && *dist < CHUNKCOPY_CHUNK_SIZE) {
storechunk(out, loadchunk(from));
out += *dist;
@@ -220,6 +225,52 @@ static inline z_vec128i_t v_load8_dup(const void* src) {
static inline void v_store_128(void* out, const z_vec128i_t vec) {
vst1q_u8(out, vec);
}
+#elif defined (INFLATE_CHUNK_SIMD_SSE2)
+/*
+ * v_load64_dup(): load *src as an unaligned 64-bit int and duplicate it in
+ * every 64-bit component of the 128-bit result (64-bit int splat).
+ */
+static inline z_vec128i_t v_load64_dup(const void* src) {
+ int64_t i64;
+ Z_BUILTIN_MEMCPY(&i64, src, sizeof(i64));
+ return _mm_set1_epi64x(i64);
+}
+
+/*
+ * v_load32_dup(): load *src as an unaligned 32-bit int and duplicate it in
+ * every 32-bit component of the 128-bit result (32-bit int splat).
+ */
+static inline z_vec128i_t v_load32_dup(const void* src) {
+ int32_t i32;
+ Z_BUILTIN_MEMCPY(&i32, src, sizeof(i32));
+ return _mm_set1_epi32(i32);
+}
+
+/*
+ * v_load16_dup(): load *src as an unaligned 16-bit int and duplicate it in
+ * every 16-bit component of the 128-bit result (16-bit int splat).
+ */
+static inline z_vec128i_t v_load16_dup(const void* src) {
+ int16_t i16;
+ Z_BUILTIN_MEMCPY(&i16, src, sizeof(i16));
+ return _mm_set1_epi16(i16);
+}
+
+/*
+ * v_load8_dup(): load the 8-bit int *src and duplicate it in every 8-bit
+ * component of the 128-bit result (8-bit int splat).
+ */
+static inline z_vec128i_t v_load8_dup(const void* src) {
+ return _mm_set1_epi8(*(const char*)src);
+}
+
+/*
+ * v_store_128(): store the 128-bit vec in a memory destination (that might
+ * not be 16-byte aligned) void* out.
+ */
+static inline void v_store_128(void* out, const z_vec128i_t vec) {
+ _mm_storeu_si128((__m128i*)out, vec);
+}
#endif
/*
@@ -228,8 +279,8 @@ static inline void v_store_128(void* out, const z_vec128i_t vec) {
* that it's OK to overwrite at least CHUNKCOPY_CHUNK_SIZE*3 bytes of output
* even if the length is shorter than this.
*/
-static inline unsigned char FAR* chunkset_core(
- unsigned char FAR* out,
+static inline unsigned char * chunkset_core(
+ unsigned char * out,
unsigned period,
unsigned len) {
z_vec128i_t v;
@@ -304,9 +355,9 @@ static inline unsigned char FAR* chunkset_core(
* This is reflected in the `restrict`-qualified pointers, allowing the
* compiler to re-order loads and stores.
*/
-static inline unsigned char FAR* chunkcopy_relaxed(
- unsigned char FAR* Z_RESTRICT out,
- const unsigned char FAR* Z_RESTRICT from,
+static inline unsigned char * chunkcopy_relaxed(
+ unsigned char * Z_RESTRICT out,
+ const unsigned char * Z_RESTRICT from,
unsigned len) {
return chunkcopy_core(out, from, len);
}
@@ -324,11 +375,11 @@ static inline unsigned char FAR* chunkcopy_relaxed(
* output buffer is beyond the end of the current copy, and this can still be
* exploited.
*/
-static inline unsigned char FAR* chunkcopy_safe(
- unsigned char FAR* out,
- const unsigned char FAR* Z_RESTRICT from,
+static inline unsigned char * chunkcopy_safe(
+ unsigned char * out,
+ const unsigned char * Z_RESTRICT from,
unsigned len,
- unsigned char FAR* limit) {
+ unsigned char * limit) {
Assert(out + len <= limit, "chunk copy exceeds safety limit");
return chunkcopy_core_safe(out, from, len, limit);
}
@@ -340,8 +391,8 @@ static inline unsigned char FAR* chunkcopy_safe(
* Assumes that len > 0 on entry, and that it's safe to write at least
* CHUNKCOPY_CHUNK_SIZE*3 bytes to the output.
*/
-static inline unsigned char FAR* chunkcopy_lapped_relaxed(
- unsigned char FAR* out,
+static inline unsigned char * chunkcopy_lapped_relaxed(
+ unsigned char * out,
unsigned dist,
unsigned len) {
if (dist < len && dist < CHUNKCOPY_CHUNK_SIZE) {
@@ -359,11 +410,11 @@ static inline unsigned char FAR* chunkcopy_lapped_relaxed(
* output buffer is beyond the end of the current copy, and this can still be
* exploited.
*/
-static inline unsigned char FAR* chunkcopy_lapped_safe(
- unsigned char FAR* out,
+static inline unsigned char * chunkcopy_lapped_safe(
+ unsigned char * out,
unsigned dist,
unsigned len,
- unsigned char FAR* limit) {
+ unsigned char * limit) {
Assert(out + len <= limit, "chunk copy exceeds safety limit");
if ((limit - out) < (ptrdiff_t)(3 * CHUNKCOPY_CHUNK_SIZE)) {
/* TODO(cavalcantii): try harder to optimise this */
@@ -394,7 +445,7 @@ typedef uint64_t inflate_holder_t;
* Ask the compiler to perform a wide, unaligned load of a uint64_t using a
* machine instruction appropriate for the uint64_t type.
*/
-static inline inflate_holder_t read64le(const unsigned char FAR *in) {
+static inline inflate_holder_t read64le(const unsigned char *in) {
inflate_holder_t input;
Z_BUILTIN_MEMCPY(&input, in, sizeof(input));
return input;
diff --git a/configure b/configure
index e094818..5f0e9bf 100755
--- a/configure
+++ b/configure
@@ -570,6 +570,32 @@ fi
rm -f conftest.c
+# Enable SIMD inflate on architectures where the selected ISA is mandatory.
+cat > conftest.c << EOF
+#if !defined(__x86_64__) && !defined(__amd64__)
+#error "not x86-64"
+#endif
+#include <emmintrin.h>
+int main(void) { return _mm_cvtsi128_si32(_mm_setzero_si128()); }
+EOF
+if check c_compile "for x86-64 SSE2 inflate chunk copy" \
+ "$CC" "$CFLAGS -msse2" "$CPPFLAGS" "$LDFLAGS" "$LIBS"; then
+ CFLAGS="$CFLAGS -msse2 -DINFLATE_CHUNK_SIMD_SSE2 -DINFLATE_CHUNK_READ_64LE"
+fi
+
+cat > conftest.c << EOF
+#if !defined(__aarch64__)
+#error "not AArch64"
+#endif
+#include <arm_neon.h>
+int main(void) { return vgetq_lane_u8(vdupq_n_u8(0), 0); }
+EOF
+if check c_compile "for AArch64 NEON inflate chunk copy" \
+ "$CC" "$CFLAGS" "$CPPFLAGS" "$LDFLAGS" "$LIBS"; then
+ CFLAGS="$CFLAGS -DINFLATE_CHUNK_SIMD_NEON -DINFLATE_CHUNK_READ_64LE -DADLER32_SIMD_NEON"
+fi
+rm -f conftest.c
+
#
# Determine which feature macros to pass.
#
diff --git a/inffast_chunk.c b/inffast_chunk.c
index d80c9a7..39eff71 100644
--- a/inffast_chunk.c
+++ b/inffast_chunk.c
@@ -95,24 +95,24 @@ void ZLIB_INTERNAL inflate_fast_chunk_(strm, start)
z_streamp strm;
unsigned start; /* inflate()'s starting value for strm->avail_out */
{
- struct inflate_state FAR *state;
- z_const unsigned char FAR *in; /* local strm->next_in */
- z_const unsigned char FAR *last; /* have enough input while in < last */
- unsigned char FAR *out; /* local strm->next_out */
- unsigned char FAR *beg; /* inflate()'s initial strm->next_out */
- unsigned char FAR *end; /* while out < end, enough space available */
- unsigned char FAR *limit; /* safety limit for chunky copies */
+ struct inflate_state *state;
+ z_const unsigned char *in; /* local strm->next_in */
+ z_const unsigned char *last; /* have enough input while in < last */
+ unsigned char *out; /* local strm->next_out */
+ unsigned char *beg; /* inflate()'s initial strm->next_out */
+ unsigned char *end; /* while out < end, enough space available */
+ unsigned char *limit; /* safety limit for chunky copies */
#ifdef INFLATE_STRICT
unsigned dmax; /* maximum distance from zlib header */
#endif
unsigned wsize; /* window size or zero if not using window */
unsigned whave; /* valid bytes in the window */
unsigned wnext; /* window write index */
- unsigned char FAR *window; /* allocated sliding window, if wsize != 0 */
+ unsigned char *window; /* allocated sliding window, if wsize != 0 */
inflate_holder_t hold; /* local strm->hold */
unsigned bits; /* local strm->bits */
- code const FAR *lcode; /* local strm->lencode */
- code const FAR *dcode; /* local strm->distcode */
+ code const *lcode; /* local strm->lencode */
+ code const *dcode; /* local strm->distcode */
unsigned lmask; /* mask for first level of length codes */
unsigned dmask; /* mask for first level of distance codes */
code here; /* retrieved table entry */
@@ -120,10 +120,10 @@ unsigned start; /* inflate()'s starting value for strm->avail_out */
/* window position, window bytes to copy */
unsigned len; /* match length, unused bytes */
unsigned dist; /* match distance */
- unsigned char FAR *from; /* where to copy match from */
+ unsigned char *from; /* where to copy match from */
/* copy state to local variables */
- state = (struct inflate_state FAR *)strm->state;
+ state = (struct inflate_state *)strm->state;
in = strm->next_in;
last = in + (strm->avail_in - (INFLATE_FAST_MIN_INPUT - 1));
out = strm->next_out;
diff --git a/inffast_chunk.h b/inffast_chunk.h
index de6aa0d..5d6025a 100644
--- a/inffast_chunk.h
+++ b/inffast_chunk.h
@@ -45,4 +45,4 @@
#define INFLATE_FAST_MIN_INPUT 8
#endif
-void ZLIB_INTERNAL inflate_fast_chunk_ OF((z_streamp strm, unsigned start));
+void ZLIB_INTERNAL inflate_fast_chunk_(z_streamp strm, unsigned start);
diff --git a/inflate.c b/inflate.c
index 626488c..e622259 100644
--- a/inflate.c
+++ b/inflate.c
@@ -83,7 +83,7 @@
#include "zutil.h"
#include "inftrees.h"
#include "inflate.h"
-#ifdef INFLATE_CHUNK_SIMD_NEON
+#if defined(INFLATE_CHUNK_SIMD_NEON) || defined(INFLATE_CHUNK_SIMD_SSE2)
#include "inffast_chunk.h"
#include "chunkcopy.h"
#else
@@ -265,7 +265,7 @@ static int updatewindow(z_stream *strm,
/* if it hasn't been done already, allocate space for the window */
if (state->window == NULL) {
-#ifdef INFLATE_CHUNK_SIMD_NEON
+#if defined(INFLATE_CHUNK_SIMD_NEON) || defined(INFLATE_CHUNK_SIMD_SSE2)
unsigned int wsize = 1U << state->wbits;
state->window = (unsigned char *)
z_stream_alloc(strm, wsize + CHUNKCOPY_CHUNK_SIZE);
@@ -891,7 +891,7 @@ int ZEXPORT inflate(z_stream *strm,
if (have >= INFLATE_FAST_MIN_INPUT &&
left >= INFLATE_FAST_MIN_OUTPUT) {
RESTORE();
-#ifdef INFLATE_CHUNK_SIMD_NEON
+#if defined(INFLATE_CHUNK_SIMD_NEON) || defined(INFLATE_CHUNK_SIMD_SSE2)
inflate_fast_chunk_(strm, out);
#else
inflate_fast(strm, out);
@@ -1011,7 +1011,7 @@ int ZEXPORT inflate(z_stream *strm,
else
from = state->window + (state->wnext - copy);
if (copy > state->length) copy = state->length;
-#ifdef INFLATE_CHUNK_SIMD_NEON
+#if defined(INFLATE_CHUNK_SIMD_NEON) || defined(INFLATE_CHUNK_SIMD_SSE2)
if (copy > left) copy = left;
put = chunkcopy_safe(put, from, copy, put + left);
}
@@ -1094,7 +1094,7 @@ int ZEXPORT inflate(z_stream *strm,
Note: a memory error from inflate() is non-recoverable.
*/
inf_leave:
-#ifdef INFLATE_CHUNK_SIMD_NEON
+#if defined(INFLATE_CHUNK_SIMD_NEON) || defined(INFLATE_CHUNK_SIMD_SSE2)
if (left >= CHUNKCOPY_CHUNK_SIZE)
memset(put, 0x55, CHUNKCOPY_CHUNK_SIZE);
else
--
2.55.0