final parameter tuning

doc/zstd_manual.html

@@ -182,6 +182,7 @@ size_t     ZSTD_freeDCtx(ZSTD_DCtx* dctx);
   ZSTD_createCDict() will create a digested dictionary, ready to start future compression operations without startup delay.
   ZSTD_CDict can be created once and shared by multiple threads concurrently, since its usage is read-only.
   `dictBuffer` can be released after ZSTD_CDict creation, since its content is copied within CDict
+  Note : A ZSTD_CDict can be created with an empty dictionary, but it is inefficient for small data. 
 </p></pre><BR>
 
 <pre><b>size_t      ZSTD_freeCDict(ZSTD_CDict* CDict);
@@ -196,6 +197,8 @@ size_t     ZSTD_freeDCtx(ZSTD_DCtx* dctx);
   Faster startup than ZSTD_compress_usingDict(), recommended when same dictionary is used multiple times.
   Note that compression level is decided during dictionary creation.
   Frame parameters are hardcoded (dictID=yes, contentSize=yes, checksum=no)
+  Note : ZSTD_compress_usingCDict() can be used with a ZSTD_CDict created from an empty dictionary.
+         But it is inefficient for small data, and it is recommended to use ZSTD_compressCCtx(). 
 </p></pre><BR>
 
 <pre><b>ZSTD_DDict* ZSTD_createDDict(const void* dictBuffer, size_t dictSize);

lib/decompress/zstd_decompress.c

@@ -1039,7 +1039,8 @@ size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
     /* prefetch dictionary content */
     if (dctx->ddictIsCold) {
         size_t const dictSize = (const char*)dctx->prefixStart - (const char*)dctx->virtualStart;
-        size_t const pSize = MIN(dictSize, (size_t)(64*nbSeq));
+        size_t const psmin = MIN(dictSize, (size_t)(64*nbSeq) /* heuristic */ );
+        size_t const pSize = MIN(psmin, 128 KB /* protection */ );
         const void* const pStart = (const char*)dctx->dictEnd - pSize;
         PREFETCH_AREA(pStart, pSize);
         dctx->ddictIsCold = 0;