minor +10% speed improvement for scalar ZSTD_get1BlockSummary()

lib/compress/zstd_compress.c

@@ -7395,14 +7395,68 @@ size_t ZSTD_convertBlockSequences(ZSTD_CCtx* cctx,
     return ZSTD_convertBlockSequences_internal(cctx, inSeqs, nbSequences, 0);
 }
 
+#if 0 && defined(__AVX2__)
+
+/* C90-compatible alignment macro (GCC/Clang). Adjust for other compilers if needed. */
+#if defined(__GNUC__)
+#  define ALIGNED32 __attribute__((aligned(32)))
+#else
+#  define ALIGNED32
+#endif
+
 BlockSummary ZSTD_get1BlockSummary(const ZSTD_Sequence* seqs, size_t nbSeqs)
 {
-    size_t blockSize = 0;
+    size_t i;
+    __m256i sumVec;            /* accumulates match+lit in 32-bit lanes */
+    __m256i mask;              /* shuffling control */
+    ALIGNED32 int tmp[8];      /* temporary buffer for reduction */
+    uint64_t sum;
+    int k;
+
+    sumVec = _mm256_setzero_si256();
+    mask   = _mm256_setr_epi32(
+                 1,5,  /* match(0), match(1) */
+                 2,6,  /* lit(0),   lit(1)   */
+                 1,5,  /* match(0), match(1) */
+                 2,6   /* lit(0),   lit(1)   */
+             );
+
+    /* Process 2 structs (32 bytes) at a time */
+    for (i = 0; i + 2 <= count; i += 2) {
+        /* Load two consecutive MyStructs (8×4 = 32 bytes) */
+        __m256i data     = _mm256_loadu_si256((const __m256i*)&arr[i]);
+        /* Shuffle out lanes 1,2,5,6 => match(0), match(1), lit(0), lit(1), repeated */
+        __m256i selected = _mm256_permutevar8x32_epi32(data, mask);
+        /* Accumulate in sumVec */
+        sumVec           = _mm256_add_epi32(sumVec, selected);
+    }
+
+    /* Horizontal reduction of sumVec */
+    _mm256_store_si256((__m256i*)tmp, sumVec);
+    sum = 0;
+    for (k = 0; k < 8; k++) {
+        sum += (uint64_t)tmp[k]; /* each lane is match+lit from pairs, repeated twice */
+    }
+
+    /* Handle the leftover (if count is odd) */
+    for (; i < count; i++) {
+        sum += arr[i].matchLength;
+        sum += arr[i].litLength;
+    }
+
+    return sum;
+}
+
+#else
+
+BlockSummary ZSTD_get1BlockSummary(const ZSTD_Sequence* seqs, size_t nbSeqs)
+{
+    size_t totalMatchSize = 0;
     size_t litSize = 0;
     size_t n;
     assert(seqs);
     for (n=0; n<nbSeqs; n++) {
-        blockSize += seqs[n].matchLength + seqs[n].litLength;
+        totalMatchSize += seqs[n].matchLength;
         litSize += seqs[n].litLength;
         if (seqs[n].matchLength == 0) {
             assert(seqs[n].offset == 0);
@@ -7416,11 +7470,12 @@ BlockSummary ZSTD_get1BlockSummary(const ZSTD_Sequence* seqs, size_t nbSeqs)
     }
     {   BlockSummary bs;
         bs.nbSequences = n+1;
-        bs.blockSize = blockSize;
+        bs.blockSize = litSize + totalMatchSize;
         bs.litSize = litSize;
         return bs;
     }
 }
+#endif
 
 
 static size_t