Compute the wordmask instead of looking it up in a table.

Tested:
  name                                old speed               new speed               delta
  BM_UFlat/0      [html   ]           2.13GB/s ± 0%           2.46GB/s ± 0%  +15.70%         (p=0.000 n=10+8)
  BM_UFlat/1      [urls   ]           1.21GB/s ± 0%           1.20GB/s ± 0%   -1.49%         (p=0.000 n=9+10)
  BM_UFlat/2      [jpg    ]           17.1GB/s ± 1%           17.2GB/s ± 1%     ~           (p=0.120 n=11+11)
  BM_UFlat/3      [jpg_200]           1.55GB/s ± 0%           1.54GB/s ± 0%   -0.96%         (p=0.000 n=10+7)
  BM_UFlat/4      [pdf    ]           12.9GB/s ± 0%           12.6GB/s ± 0%   -1.98%         (p=0.000 n=11+9)
  BM_UFlat/5      [html4  ]           1.87GB/s ± 0%           1.87GB/s ± 0%   -0.06%        (p=0.033 n=11+11)
  BM_UFlat/6      [txt1   ]            816MB/s ± 0%            793MB/s ± 0%   -2.84%        (p=0.000 n=11+11)
  BM_UFlat/7      [txt2   ]            758MB/s ± 0%            737MB/s ± 0%   -2.77%        (p=0.000 n=11+11)
  BM_UFlat/8      [txt3   ]            865MB/s ± 0%            839MB/s ± 0%   -2.94%         (p=0.000 n=11+8)
  BM_UFlat/9      [txt4   ]            701MB/s ± 0%            679MB/s ± 0%   -3.11%        (p=0.000 n=11+10)
  BM_UFlat/10     [pb     ]           2.60GB/s ± 2%           3.07GB/s ± 0%  +17.81%        (p=0.000 n=11+11)
  BM_UFlat/11     [gaviota]           1.01GB/s ± 0%           0.97GB/s ± 0%   -3.83%        (p=0.000 n=11+10)
  BM_UFlat/12     [cp     ]           1.66GB/s ± 1%           1.73GB/s ± 1%   +4.32%        (p=0.000 n=11+11)
  BM_UFlat/13     [c      ]           1.52GB/s ± 1%           1.53GB/s ± 0%   +0.49%        (p=0.002 n=11+11)
  BM_UFlat/14     [lsp    ]           1.61GB/s ± 0%           1.64GB/s ± 0%   +2.10%        (p=0.000 n=10+11)
  BM_UFlat/15     [xls    ]           1.12GB/s ± 0%           1.08GB/s ± 0%   -3.95%         (p=0.000 n=11+7)
  BM_UFlat/16     [xls_200]            926MB/s ± 1%            935MB/s ± 1%     ~            (p=0.056 n=9+11)
  BM_UFlat/17     [bin    ]           1.89GB/s ± 0%           1.86GB/s ± 0%   -1.32%        (p=0.000 n=11+11)
  BM_UFlat/18     [bin_200]           1.96GB/s ± 0%           1.99GB/s ± 1%   +1.78%        (p=0.000 n=11+11)
  BM_UFlat/19     [sum    ]           1.32GB/s ± 0%           1.31GB/s ± 0%   -0.79%        (p=0.000 n=11+10)
  BM_UFlat/20     [man    ]           1.40GB/s ± 0%           1.43GB/s ± 0%   +2.51%         (p=0.000 n=9+10)
  BM_UValidate/0  [html   ]           2.95GB/s ± 1%           3.07GB/s ± 0%   +4.11%        (p=0.000 n=10+11)
  BM_UValidate/1  [urls   ]           1.57GB/s ± 0%           1.60GB/s ± 0%   +2.24%        (p=0.000 n=10+11)
  BM_UValidate/2  [jpg    ]            822GB/s ± 0%            850GB/s ± 0%   +3.42%        (p=0.000 n=10+11)
  BM_UValidate/3  [jpg_200]           2.01GB/s ± 0%           2.04GB/s ± 0%   +1.24%        (p=0.000 n=11+11)
  BM_UValidate/4  [pdf    ]           33.7GB/s ± 0%           35.9GB/s ± 1%   +6.51%        (p=0.000 n=10+11)
  BM_UIOVec/0     [html   ]            852MB/s ± 0%            852MB/s ± 0%     ~           (p=0.898 n=11+11)
  BM_UIOVec/1     [urls   ]            663MB/s ± 0%            652MB/s ± 0%   -1.61%        (p=0.000 n=11+11)
  BM_UIOVec/2     [jpg    ]           15.3GB/s ± 1%           15.3GB/s ± 2%     ~            (p=0.459 n=9+10)
  BM_UIOVec/3     [jpg_200]            652MB/s ± 0%            627MB/s ± 1%   -3.80%        (p=0.000 n=10+11)
  BM_UIOVec/4     [pdf    ]           8.80GB/s ± 1%           8.57GB/s ± 1%   -2.62%        (p=0.000 n=10+11)
  BM_UFlatSink/0  [html   ]           2.13GB/s ± 0%           2.46GB/s ± 0%  +15.63%        (p=0.000 n=11+11)
  BM_UFlatSink/1  [urls   ]           1.21GB/s ± 0%           1.20GB/s ± 0%   -1.42%        (p=0.000 n=11+10)
  BM_UFlatSink/2  [jpg    ]           17.1GB/s ± 2%           17.2GB/s ± 1%     ~            (p=0.175 n=11+9)
  BM_UFlatSink/3  [jpg_200]           1.52GB/s ± 1%           1.47GB/s ± 3%   -3.15%        (p=0.000 n=11+11)
  BM_UFlatSink/4  [pdf    ]           12.8GB/s ± 1%           12.6GB/s ± 1%   -1.76%        (p=0.000 n=11+11)
  BM_UFlatSink/5  [html4  ]           1.87GB/s ± 0%           1.87GB/s ± 0%   -0.19%        (p=0.000 n=11+10)
  BM_UFlatSink/6  [txt1   ]            816MB/s ± 0%            792MB/s ± 0%   -2.94%        (p=0.000 n=11+11)
  BM_UFlatSink/7  [txt2   ]            758MB/s ± 0%            736MB/s ± 0%   -2.83%        (p=0.000 n=11+11)
  BM_UFlatSink/8  [txt3   ]            865MB/s ± 0%            838MB/s ± 0%   -3.13%        (p=0.000 n=11+11)
  BM_UFlatSink/9  [txt4   ]            701MB/s ± 0%            678MB/s ± 0%   -3.20%        (p=0.000 n=11+11)
  BM_UFlatSink/10 [pb     ]           2.60GB/s ± 2%           3.07GB/s ± 0%  +18.27%        (p=0.000 n=11+10)
  BM_UFlatSink/11 [gaviota]           1.01GB/s ± 0%           0.97GB/s ± 0%   -3.90%        (p=0.000 n=11+11)
  BM_UFlatSink/12 [cp     ]           1.66GB/s ± 1%           1.73GB/s ± 1%   +4.62%        (p=0.000 n=11+10)
  BM_UFlatSink/13 [c      ]           1.52GB/s ± 0%           1.53GB/s ± 1%     ~            (p=0.180 n=9+11)
  BM_UFlatSink/14 [lsp    ]           1.61GB/s ± 0%           1.64GB/s ± 1%   +1.98%         (p=0.000 n=9+11)
  BM_UFlatSink/15 [xls    ]           1.12GB/s ± 0%           1.08GB/s ± 0%   -3.76%        (p=0.000 n=11+11)
  BM_UFlatSink/16 [xls_200]            909MB/s ± 2%            924MB/s ± 1%   +1.62%        (p=0.000 n=11+11)
  BM_UFlatSink/17 [bin    ]           1.88GB/s ± 0%           1.86GB/s ± 0%   -1.18%         (p=0.000 n=9+11)
  BM_UFlatSink/18 [bin_200]           1.94GB/s ± 2%           1.94GB/s ± 1%     ~           (p=0.090 n=11+11)
  BM_UFlatSink/19 [sum    ]           1.32GB/s ± 0%           1.31GB/s ± 0%   -0.76%        (p=0.000 n=11+11)
  BM_UFlatSink/20 [man    ]           1.39GB/s ± 2%           1.43GB/s ± 0%   +2.75%        (p=0.000 n=11+10)

  Assembly before:

*	44 8b 5c 85 a0       	mov    -0x60(%rbp,%rax,4),%r11d
	45 23 5d 00          	and    0x0(%r13),%r11d
	89 d6                	mov    %edx,%esi
	81 e6 00 07 00 00    	and    $0x700,%esi

  Assembly after:

*	89 c1                	mov    %eax,%ecx
*	c0 e1 03             	shl    $0x3,%cl
*	bf ff ff ff ff       	mov    $0xffffffff,%edi
*	48 d3 e7             	shl    %cl,%rdi
*	f7 d7                	not    %edi
	41 23 7d 00          	and    0x0(%r13),%edi
	41 89 d3             	mov    %edx,%r11d
	41 81 e3 00 07 00 00 	and    $0x700,%r11d

1 files changed, 11 insertions, 24 deletions

diff --git a/snappy.cc b/snappy.cc
index e594bb9..5352c24 100644
--- a/snappy.cc
+++ b/snappy.cc
@@ -675,14 +675,15 @@ static inline void Report(const char *algorithm, size_t compressed_size,
 //   bool TryFastAppend(const char* ip, size_t available, size_t length);
 // };
 
-namespace internal {
-
-// Mapping from i in range [0,4] to a mask to extract the bottom 8*i bits
-static const uint32 wordmask[] = {
-  0u, 0xffu, 0xffffu, 0xffffffu, 0xffffffffu
-};
-
-}  // end namespace internal
+// Mapping from n in range [0,4] to a mask to extract the bottom 8*n bits.
+static inline uint32 WordMask(int n) {
+  DCHECK_GE(n, 0);
+  DCHECK_LE(n, 4);
+  // This needs to be wider than uint32 otherwise `mask << 32` will be
+  // undefined.
+  uint64 mask = 0xffffffff;
+  return ~(mask << (8 * n));
+}
 
 // Helper class for decompression
 class SnappyDecompressor {
@@ -770,20 +771,6 @@ class SnappyDecompressor {
 #endif
 
     const char* ip = ip_;
-    // For position-independent executables, accessing global arrays can be
-    // slow.  Move wordmask array onto the stack to mitigate this.
-    uint32 wordmask[sizeof(internal::wordmask)/sizeof(uint32)];
-    // Do not use memcpy to copy internal::wordmask to
-    // wordmask.  LLVM converts stack arrays to global arrays if it detects
-    // const stack arrays and this hurts the performance of position
-    // independent code. This change is temporary and can be reverted when
-    // https://reviews.llvm.org/D30759 is approved.
-    wordmask[0] = internal::wordmask[0];
-    wordmask[1] = internal::wordmask[1];
-    wordmask[2] = internal::wordmask[2];
-    wordmask[3] = internal::wordmask[3];
-    wordmask[4] = internal::wordmask[4];
-
     // We could have put this refill fragment only at the beginning of the loop.
     // However, duplicating it at the end of each branch gives the compiler more
     // scope to optimize the <ip_limit_ - ip> expression based on the local
@@ -825,7 +812,7 @@ class SnappyDecompressor {
           // Long literal.
           const size_t literal_length_length = literal_length - 60;
           literal_length =
-              (LittleEndian::Load32(ip) & wordmask[literal_length_length]) + 1;
+              (LittleEndian::Load32(ip) & WordMask(literal_length_length)) + 1;
           ip += literal_length_length;
         }
 
@@ -848,7 +835,7 @@ class SnappyDecompressor {
         MAYBE_REFILL();
       } else {
         const size_t entry = char_table[c];
-        const size_t trailer = LittleEndian::Load32(ip) & wordmask[entry >> 11];
+        const size_t trailer = LittleEndian::Load32(ip) & WordMask(entry >> 11);
         const size_t length = entry & 0xff;
         ip += entry >> 11;