aboutsummaryrefslogtreecommitdiff
path: root/port/port_posix_sse.cc
diff options
context:
space:
mode:
authorPieter Wuille <pieter.wuille@gmail.com>2017-06-09 19:24:30 -0700
committerPieter Wuille <pieter.wuille@gmail.com>2017-06-09 19:24:30 -0700
commitcf44e4ca7762742c6c3154447b40869ec9d041db (patch)
treed5a89851da0a8ab07c40966ac137f4e299c80927 /port/port_posix_sse.cc
parent634ad517037b319147816f1d112b066528e1724a (diff)
downloadbitcoin-cf44e4ca7762742c6c3154447b40869ec9d041db.tar.xz
Squashed 'src/leveldb/' changes from a31c8aa40..196962ff0
196962ff0 Add AcceleratedCRC32C to port_win.h 1bdf1c34c Merge upstream LevelDB v1.20 d31721eb0 Merge #17: Fixed file sharing errors fecd44902 Fixed file sharing error in Win32Env::GetFileSize(), Win32SequentialFile::_Init(), Win32RandomAccessFile::_Init() Fixed error checking in Win32SequentialFile::_Init() 5b7510f1b Merge #14: Merge upstream LevelDB 1.19 0d969fd57 Merge #16: [LevelDB] Do no crash if filesystem can't fsync c8c029b5b [LevelDB] Do no crash if filesystem can't fsync a53934a3a Increase leveldb version to 1.20. f3f139737 Separate Env tests from PosixEnv tests. eb4f0972f leveldb: Fix compilation warnings in port_posix_sse.cc on x86 (32-bit). d0883b600 Fixed path to doc file: index.md. 7fa20948d Convert documentation to markdown. ea175e28f Implement support for Intel crc32 instruction (SSE 4.2) 95cd743e5 Including <limits> for std::numeric_limits. 646c3588d Limit the number of read-only files the POSIX Env will have open. d40bc3fa5 Merge #13: Typo ebbd772d3 Typo a2fb086d0 Add option for max file size. The currend hard-coded value of 2M is inefficient in colossus. git-subtree-dir: src/leveldb git-subtree-split: 196962ff01c39b4705d8117df5c3f8c205349950
Diffstat (limited to 'port/port_posix_sse.cc')
-rw-r--r--port/port_posix_sse.cc129
1 files changed, 129 insertions, 0 deletions
diff --git a/port/port_posix_sse.cc b/port/port_posix_sse.cc
new file mode 100644
index 0000000000..1e519ba0b6
--- /dev/null
+++ b/port/port_posix_sse.cc
@@ -0,0 +1,129 @@
+// Copyright 2016 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+//
+// A portable implementation of crc32c, optimized to handle
+// four bytes at a time.
+//
+// In a separate source file to allow this accelerated CRC32C function to be
+// compiled with the appropriate compiler flags to enable x86 SSE 4.2
+// instructions.
+
+#include <stdint.h>
+#include <string.h>
+#include "port/port.h"
+
+#if defined(LEVELDB_PLATFORM_POSIX_SSE)
+
+#if defined(_MSC_VER)
+#include <intrin.h>
+#elif defined(__GNUC__) && defined(__SSE4_2__)
+#include <nmmintrin.h>
+#include <cpuid.h>
+#endif
+
+#endif // defined(LEVELDB_PLATFORM_POSIX_SSE)
+
+namespace leveldb {
+namespace port {
+
+#if defined(LEVELDB_PLATFORM_POSIX_SSE)
+
+// Used to fetch a naturally-aligned 32-bit word in little endian byte-order
+static inline uint32_t LE_LOAD32(const uint8_t *p) {
+ // SSE is x86 only, so ensured that |p| is always little-endian.
+ uint32_t word;
+ memcpy(&word, p, sizeof(word));
+ return word;
+}
+
+#if defined(_M_X64) || defined(__x86_64__) // LE_LOAD64 is only used on x64.
+
+// Used to fetch a naturally-aligned 64-bit word in little endian byte-order
+static inline uint64_t LE_LOAD64(const uint8_t *p) {
+ uint64_t dword;
+ memcpy(&dword, p, sizeof(dword));
+ return dword;
+}
+
+#endif // defined(_M_X64) || defined(__x86_64__)
+
+static inline bool HaveSSE42() {
+#if defined(_MSC_VER)
+ int cpu_info[4];
+ __cpuid(cpu_info, 1);
+ return (cpu_info[2] & (1 << 20)) != 0;
+#elif defined(__GNUC__)
+ unsigned int eax, ebx, ecx, edx;
+ __get_cpuid(1, &eax, &ebx, &ecx, &edx);
+ return (ecx & (1 << 20)) != 0;
+#else
+ return false;
+#endif
+}
+
+#endif // defined(LEVELDB_PLATFORM_POSIX_SSE)
+
+// For further improvements see Intel publication at:
+// http://download.intel.com/design/intarch/papers/323405.pdf
+uint32_t AcceleratedCRC32C(uint32_t crc, const char* buf, size_t size) {
+#if !defined(LEVELDB_PLATFORM_POSIX_SSE)
+ return 0;
+#else
+ static bool have = HaveSSE42();
+ if (!have) {
+ return 0;
+ }
+
+ const uint8_t *p = reinterpret_cast<const uint8_t *>(buf);
+ const uint8_t *e = p + size;
+ uint32_t l = crc ^ 0xffffffffu;
+
+#define STEP1 do { \
+ l = _mm_crc32_u8(l, *p++); \
+} while (0)
+#define STEP4 do { \
+ l = _mm_crc32_u32(l, LE_LOAD32(p)); \
+ p += 4; \
+} while (0)
+#define STEP8 do { \
+ l = _mm_crc32_u64(l, LE_LOAD64(p)); \
+ p += 8; \
+} while (0)
+
+ if (size > 16) {
+ // Process unaligned bytes
+ for (unsigned int i = reinterpret_cast<uintptr_t>(p) % 8; i; --i) {
+ STEP1;
+ }
+
+ // _mm_crc32_u64 is only available on x64.
+#if defined(_M_X64) || defined(__x86_64__)
+ // Process 8 bytes at a time
+ while ((e-p) >= 8) {
+ STEP8;
+ }
+ // Process 4 bytes at a time
+ if ((e-p) >= 4) {
+ STEP4;
+ }
+#else // !(defined(_M_X64) || defined(__x86_64__))
+ // Process 4 bytes at a time
+ while ((e-p) >= 4) {
+ STEP4;
+ }
+#endif // defined(_M_X64) || defined(__x86_64__)
+ }
+ // Process the last few bytes
+ while (p != e) {
+ STEP1;
+ }
+#undef STEP8
+#undef STEP4
+#undef STEP1
+ return l ^ 0xffffffffu;
+#endif // defined(LEVELDB_PLATFORM_POSIX_SSE)
+}
+
+} // namespace port
+} // namespace leveldb