From: Yann Collet <cyan@fb.com>
Date: Mon, 5 Feb 2018 19:46:02 +0000 (-0800)
Subject: minor : modified ZSTD_preserveUnsortedMark() to be more vectorization friendly
X-Git-Tag: v1.3.4~1^2~61^2~1
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=0170cf9a7af94d9ffdcdb5cd3ebc35fcd542f2ed;p=thirdparty%2Fzstd.git

minor : modified ZSTD_preserveUnsortedMark() to be more vectorization friendly
---

diff --git a/lib/common/fse_decompress.c b/lib/common/fse_decompress.c
index 8e3f0035f..4c66c3b77 100644
--- a/lib/common/fse_decompress.c
+++ b/lib/common/fse_decompress.c
@@ -139,8 +139,8 @@ size_t FSE_buildDTable(FSE_DTable* dt, const short* normalizedCounter, unsigned
     {   U32 u;
         for (u=0; u<tableSize; u++) {
             FSE_FUNCTION_TYPE const symbol = (FSE_FUNCTION_TYPE)(tableDecode[u].symbol);
-            U16 nextState = symbolNext[symbol]++;
-            tableDecode[u].nbBits = (BYTE) (tableLog - BIT_highbit32 ((U32)nextState) );
+            U32 const nextState = symbolNext[symbol]++;
+            tableDecode[u].nbBits = (BYTE) (tableLog - BIT_highbit32(nextState) );
             tableDecode[u].newState = (U16) ( (nextState << tableDecode[u].nbBits) - tableSize);
     }   }
 
diff --git a/lib/compress/zstd_lazy.c b/lib/compress/zstd_lazy.c
index 7f59eb34a..844afc051 100644
--- a/lib/compress/zstd_lazy.c
+++ b/lib/compress/zstd_lazy.c
@@ -18,33 +18,46 @@
 #define ZSTD_DUBT_UNSORTED_MARK 1   /* note : index 1 will now be confused with "unsorted" if sorted as larger than its predecessor.
                                        It's not a big deal though : the candidate will just be considered unsorted, and be sorted again.
                                        Additionnally, candidate position 1 will be lost.
-                                       But candidate 1 cannot hide a large tree of candidates, so it's a moderate loss.
-                                       The benefit is that ZSTD_DUBT_UNSORTED_MARK cannot be misdhandled by a table re-use using a different strategy */
+                                       But candidate 1 cannot hide a large tree of candidates, so it's a minimal loss.
+                                       The benefit is that ZSTD_DUBT_UNSORTED_MARK cannot be misdhandled after table re-use with a different strategy */
 
 /*! ZSTD_preserveUnsortedMark() :
  *  pre-emptively increase value of ZSTD_DUBT_UNSORTED_MARK before ZSTD_reduceTable()
  *  so that combined operation preserves its value.
  *  Without it, ZSTD_DUBT_UNSORTED_MARK==1 would be squashed to 0.
- *  As a consequence, the list of unsorted elements would stop on the first element,
- *  removing candidates, resulting in a negligible loss to compression ratio
+ *  As a consequence, the list of unsorted elements would stop at first element,
+ *  removing candidates, resulting in a very small loss to compression ratio
  *  (since overflow protection with ZSTD_reduceTable() is relatively rare).
+ *
  *  Another potential risk is that a position will be promoted from *unsorted*
- *  to *sorted=>smaller:0*, meaning the next candidate will be considered smaller.
+ *  to *sorted=>smaller:0*, meaning next candidate will be considered smaller.
  *  This could be wrong, and result in data corruption.
+ *
  *  On second thought, this corruption might be impossible,
- *  because unsorted elements are always at the beginning of the list,
- *  and squashing to zero reduce the list to a single element,
+ *  because unsorted elements stand at the beginning of the list,
+ *  and squashing to zero reduces the list to a single element,
  *  which needs to be sorted anyway.
  *  I haven't spent much thoughts into this possible scenario,
- *  and just felt it was safer to implement ZSTD_preserveUnsortedMark() */
+ *  and just felt it was safer to implement ZSTD_preserveUnsortedMark()
+ *
+ * `size` : must be a positive multiple of ZSTD_ROWSIZE */
+#define ZSTD_ROWSIZE 16
 void ZSTD_preserveUnsortedMark (U32* const table, U32 const size, U32 const reducerValue)
 {
-    U32 u;
-    for (u=0; u<size; u++)
-        if (table[u] == ZSTD_DUBT_UNSORTED_MARK)
-            table[u] = ZSTD_DUBT_UNSORTED_MARK + reducerValue;
+    int cellNb = 0;
+    U32 const nbRows = size / ZSTD_ROWSIZE;
+    U32 rowNb;
+    assert((size % ZSTD_ROWSIZE) == 0);
+    for (rowNb=0 ; rowNb < nbRows ; rowNb++) {
+        int column;
+        for (column=0; column<ZSTD_ROWSIZE; column++) {
+            U32 const adder = (table[cellNb] == ZSTD_DUBT_UNSORTED_MARK) ? reducerValue : 0;
+            table[cellNb] += adder;
+            cellNb++;
+    }   }
 }
 
+
 void ZSTD_updateDUBT(
                 ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
                 const BYTE* ip, const BYTE* iend,
diff --git a/lib/decompress/zstd_decompress.c b/lib/decompress/zstd_decompress.c
index 1a1fb2506..afae80d1c 100644
--- a/lib/decompress/zstd_decompress.c
+++ b/lib/decompress/zstd_decompress.c
@@ -655,6 +655,7 @@ size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
 
 typedef union {
     FSE_decode_t realData;
+    FSE_DTable dtable;
     U32 alignedBy4;
 } FSE_decode_t4;
 
@@ -733,7 +734,6 @@ static size_t ZSTD_buildSeqTable(FSE_DTable* DTableSpace, const FSE_DTable** DTa
                                  const void* src, size_t srcSize,
                                  const FSE_decode_t4* defaultTable, U32 flagRepeatTable)
 {
-    const void* const tmpPtr = defaultTable;   /* bypass strict aliasing */
     switch(type)
     {
     case set_rle :
@@ -743,7 +743,7 @@ static size_t ZSTD_buildSeqTable(FSE_DTable* DTableSpace, const FSE_DTable** DTa
         *DTablePtr = DTableSpace;
         return 1;
     case set_basic :
-        *DTablePtr = (const FSE_DTable*)tmpPtr;
+        *DTablePtr = &defaultTable->dtable;
         return 0;
     case set_repeat:
         if (!flagRepeatTable) return ERROR(corruption_detected);