X-Git-Url: http://plrg.eecs.uci.edu/git/?p=folly.git;a=blobdiff_plain;f=folly%2FGroupVarint.h;h=0e8f5edd74dc89544f54e3385cb1ffbd24e62b21;hp=dc26919547da4f9fe5217408ef2c1e3ed9849a04;hb=43d3a315bbd54cb187021897b566ffb53290a44b;hpb=5d43d2ab506d7b02a54ef5a5e1d662a1911d1aa2

diff --git a/folly/GroupVarint.h b/folly/GroupVarint.h
index dc269195..0e8f5edd 100644
--- a/folly/GroupVarint.h
+++ b/folly/GroupVarint.h
@@ -1,5 +1,5 @@
 /*
- * Copyright 2013 Facebook, Inc.
+ * Copyright 2016 Facebook, Inc.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -14,28 +14,30 @@
  * limitations under the License.
  */
 
-#ifndef FOLLY_GROUPVARINT_H_
-#define FOLLY_GROUPVARINT_H_
+#pragma once
 
-#ifndef __GNUC__
-#error GroupVarint.h requires GCC
+#if !defined(__GNUC__) && !defined(_MSC_VER)
+#error GroupVarint.h requires GCC or MSVC
 #endif
 
-#if defined(__x86_64__) || defined(__i386__)
+#include <folly/Portability.h>
+
+#if FOLLY_X64 || defined(__i386__) || FOLLY_PPC64 || FOLLY_A64
 #define HAVE_GROUP_VARINT 1
 
 #include <cstdint>
 #include <limits>
-#include "folly/detail/GroupVarintDetail.h"
-#include "folly/Bits.h"
-#include "folly/Range.h"
+#include <folly/detail/GroupVarintDetail.h>
+#include <folly/Bits.h>
+#include <folly/Range.h>
+#include <folly/portability/Builtins.h>
 #include <glog/logging.h>
 
-#ifdef __SSSE3__
-#include <x86intrin.h>
+#if FOLLY_SSE >= 3
+#include <nmmintrin.h>
 namespace folly {
 namespace detail {
-extern const __m128i groupVarintSSEMasks[];
+alignas(16) extern const uint64_t groupVarintSSEMasks[];
 }  // namespace detail
 }  // namespace folly
 #endif
@@ -174,7 +176,7 @@ class GroupVarint<uint32_t> : public detail::GroupVarintBase<uint32_t> {
     p += k2+1;
     size_t k3 = b3key(k);
     *d = loadUnaligned<uint32_t>(p) & kMask[k3];
-    p += k3+1;
+    // p += k3+1;
     return end;
   }
 
@@ -186,25 +188,35 @@ class GroupVarint<uint32_t> : public detail::GroupVarintBase<uint32_t> {
     return decode_simple(p, dest, dest+1, dest+2, dest+3);
   }
 
-#ifdef __SSSE3__
+#if FOLLY_SSE >= 3
+  /**
+   * Just like the non-SSSE3 decode below, but with the additional constraint
+   * that we must be able to read at least 17 bytes from the input pointer, p.
+   */
   static const char* decode(const char* p, uint32_t* dest) {
     uint8_t key = p[0];
     __m128i val = _mm_loadu_si128((const __m128i*)(p+1));
-    __m128i mask = detail::groupVarintSSEMasks[key];
+    __m128i mask =
+        _mm_load_si128((const __m128i*)&detail::groupVarintSSEMasks[key * 2]);
     __m128i r = _mm_shuffle_epi8(val, mask);
     _mm_storeu_si128((__m128i*)dest, r);
     return p + detail::groupVarintLengths[key];
   }
 
+  /**
+   * Just like decode_simple, but with the additional constraint that
+   * we must be able to read at least 17 bytes from the input pointer, p.
+   */
   static const char* decode(const char* p, uint32_t* a, uint32_t* b,
                             uint32_t* c, uint32_t* d) {
     uint8_t key = p[0];
     __m128i val = _mm_loadu_si128((const __m128i*)(p+1));
-    __m128i mask = detail::groupVarintSSEMasks[key];
+    __m128i mask =
+        _mm_load_si128((const __m128i*)&detail::groupVarintSSEMasks[key * 2]);
     __m128i r = _mm_shuffle_epi8(val, mask);
 
     // Extracting 32 bits at a time out of an XMM register is a SSE4 feature
-#ifdef __SSE4__
+#if FOLLY_SSE >= 4
     *a = _mm_extract_epi32(r, 0);
     *b = _mm_extract_epi32(r, 1);
     *c = _mm_extract_epi32(r, 2);
@@ -502,22 +514,24 @@ class GroupVarintDecoder {
   typedef GroupVarint<T> Base;
   typedef T type;
 
-  GroupVarintDecoder() { }
+  GroupVarintDecoder() = default;
 
   explicit GroupVarintDecoder(StringPiece data,
                               size_t maxCount = (size_t)-1)
     : rrest_(data.end()),
       p_(data.data()),
       end_(data.end()),
+      limit_(end_),
       pos_(0),
       count_(0),
       remaining_(maxCount) {
   }
 
-  void reset(StringPiece data, size_t maxCount=(size_t)-1) {
+  void reset(StringPiece data, size_t maxCount = (size_t)-1) {
     rrest_ = data.end();
     p_ = data.data();
     end_ = data.end();
+    limit_ = end_;
     pos_ = 0;
     count_ = 0;
     remaining_ = maxCount;
@@ -540,10 +554,11 @@ class GroupVarintDecoder {
       // The best way to ensure this is to ensure that data has at least
       // Base::kMaxSize - 1 bytes readable *after* the end, otherwise we'll copy
       // into a temporary buffer.
-      if (rem < Base::kMaxSize) {
+      if (limit_ - p_ < Base::kMaxSize) {
         memcpy(tmp_, p_, rem);
         p_ = tmp_;
         end_ = p_ + rem;
+        limit_ = tmp_ + sizeof(tmp_);
       }
       pos_ = 0;
       const char* n = Base::decode(p_, buf_);
@@ -591,7 +606,8 @@ class GroupVarintDecoder {
   const char* rrest_;
   const char* p_;
   const char* end_;
-  char tmp_[Base::kMaxSize];
+  const char* limit_;
+  char tmp_[2 * Base::kMaxSize];
   type buf_[Base::kGroupSize];
   size_t pos_;
   size_t count_;
@@ -603,6 +619,4 @@ typedef GroupVarintDecoder<uint64_t> GroupVarint64Decoder;
 
 }  // namespace folly
 
-#endif /* defined(__x86_64__) || defined(__i386__) */
-#endif /* FOLLY_GROUPVARINT_H_ */
-
+#endif /* FOLLY_X64 || defined(__i386__) || FOLLY_PPC64 */