Under ARM, it is slightly better to reverse the word once and then extract the bits. (#1545)

* Under ARM, it is slightly better to reverse the word once and then extract the bits.

* Guarding the zero_leading_bit call to avoid sanitizer warnings.

include/simdjson/arm64/bitmanipulation.h

@@ -46,6 +46,40 @@ simdjson_really_inline int count_ones(uint64_t input_num) {
    return vaddv_u8(vcnt_u8(vcreate_u8(input_num)));
 }
 
+
+#if defined(__GNUC__) // catches clang and gcc
+/**
+ * ARM has a fast 64-bit "bit reversal function" that is handy. However,
+ * it is not generally available as an intrinsic function under Visual
+ * Studio (though this might be changing). Even under clang/gcc, we
+ * apparently need to invoke inline assembly.
+ */
+/*
+ * We use SIMDJSON_PREFER_REVERSE_BITS as a hint that algorithms that
+ * work well with bit reversal may use it.
+ */
+#define SIMDJSON_PREFER_REVERSE_BITS 1
+
+/* reverse the bits */
+simdjson_really_inline uint64_t reverse_bits(uint64_t input_num) {
+  uint64_t rev_bits;
+  __asm("rbit %0, %1" : "=r"(rev_bits) : "r"(input_num));
+  return rev_bits;
+}
+
+/**
+ * Flips bit at index 63 - lz. Thus if you have 'leading_zeroes' leading zeroes,
+ * then this will set to zero the leading bit. It is possible for leading_zeroes to be
+ * greating or equal to 63 in which case we trigger undefined behavior, but the ouput
+ * of such undefined behavior is never used.
+ **/
+NO_SANITIZE_UNDEFINED
+simdjson_really_inline uint64_t zero_leading_bit(uint64_t rev_bits, int leading_zeroes) {
+  return rev_bits ^ (uint64_t(0x8000000000000000) >> leading_zeroes);
+}
+
+#endif
+
 simdjson_really_inline bool add_overflow(uint64_t value1, uint64_t value2, uint64_t *result) {
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
   *result = value1 + value2;

src/generic/stage1/json_structural_indexer.h

@@ -31,8 +31,62 @@ public:
     // it helps tremendously.
     if (bits == 0)
         return;
-    int cnt = static_cast<int>(count_ones(bits));
+#if defined(SIMDJSON_PREFER_REVERSE_BITS)
+    /**
+     * ARM lacks a fast trailing zero instruction, but it has a fast
+     * bit reversal instruction and a fast leading zero instruction.
+     * Thus it may be profitable to reverse the bits (once) and then
+     * to rely on a sequence of instructions that call the leading
+     * zero instruction.
+     *
+     * Performance notes:
+     * The chosen routine is not optimal in terms of data dependency
+     * since zero_leading_bit might require two instructions. However,
+     * it tends to minimize the total number of instructions which is
+     * beneficial.
+     */
 
+    uint64_t rev_bits = reverse_bits(bits);
+    int cnt = static_cast<int>(count_ones(bits));
+    int i = 0;
+    // Do the first 8 all together
+    for (; i<8; i++) {
+      int lz = leading_zeroes(rev_bits);
+      this->tail[i] = static_cast<uint32_t>(idx) + lz;
+      rev_bits = zero_leading_bit(rev_bits, lz);
+    }
+    // Do the next 8 all together (we hope in most cases it won't happen at all
+    // and the branch is easily predicted).
+    if (simdjson_unlikely(cnt > 8)) {
+      i = 8;
+      for (; i<16; i++) {
+        int lz = leading_zeroes(rev_bits);
+        this->tail[i] = static_cast<uint32_t>(idx) + lz;
+        rev_bits = zero_leading_bit(rev_bits, lz);
+      }
+
+
+      // Most files don't have 16+ structurals per block, so we take several basically guaranteed
+      // branch mispredictions here. 16+ structurals per block means either punctuation ({} [] , :)
+      // or the start of a value ("abc" true 123) every four characters.
+      if (simdjson_unlikely(cnt > 16)) {
+        i = 16;
+        while (rev_bits != 0) {
+          int lz = leading_zeroes(rev_bits);
+          this->tail[i++] = static_cast<uint32_t>(idx) + lz;
+          rev_bits = zero_leading_bit(rev_bits, lz);
+        }
+      }
+    }
+    this->tail += cnt;
+#else // SIMDJSON_PREFER_REVERSE_BITS
+    /**
+     * Under recent x64 systems, we often have both a fast trailing zero
+     * instruction and a fast 'clear-lower-bit' instruction so the following
+     * algorithm can be competitive.
+     */
+
+    int cnt = static_cast<int>(count_ones(bits));
     // Do the first 8 all together
     for (int i=0; i<8; i++) {
       this->tail[i] = idx + trailing_zeroes(bits);
@@ -61,6 +115,7 @@ public:
     }
 
     this->tail += cnt;
+#endif
   }
 };