feat: add loop-unroll support

src/core/ranges/internal/derive.cc

@@ -39,7 +39,7 @@ void Ranges::derive(const int head, Ranges &output) const {
             /// ( xx xx xx ) xx xx xx ... [reversed range]
             ///         +1   00 00 00 ...     (delta)
             const uint32_t delta = 1U << (32 - offset * 2); // distance to next possible range
-            const auto min_next = delta + range_reverse((*this)[index]) & ~(delta - 1);
+            const auto min_next = delta + (range_reverse((*this)[index]) & ~(delta - 1));
             if (min_next > max_val) {
                 break; // index has overflowed
             }

src/core/ranges/internal/ranges.cc

@@ -39,7 +39,7 @@ KLSK_INLINE CommonCode RangesUnion::operator[](size_type n) const {
     }
     n -= ranges(0).size();
 
-    #pragma unroll
+    KLSK_UNROLL(sizeof(Heads) - 2)
     for (const uint64_t head : std::to_array({0x1, 0x2, 0x4, 0x5, 0x6, 0x8, 0x9, 0xA, 0xC, 0xD})) {
         if (n < ranges(head).size()) {
             return CommonCode::unsafe_create(head << 32 | ranges(head)[n]);
@@ -48,7 +48,4 @@ KLSK_INLINE CommonCode RangesUnion::operator[](size_type n) const {
     }
 
     return CommonCode::unsafe_create((uint64_t)0xE << 32 | ranges(0xE)[n]);
-
-    // std::unreachable();
-
 }

src/core/ranges/internal/ranges_union.inl

@@ -12,7 +12,7 @@ inline const Ranges& RangesUnion::ranges(const size_t head) const {
 
 KLSK_INLINE_H size_t RangesUnion::size() const {
     size_type size = 0;
-    #pragma unroll
+    KLSK_UNROLL(sizeof(Heads))
     for (const auto head : Heads) {
         size += ranges(head).size();
     }

src/core/utils/utility.h

@@ -41,7 +41,15 @@
 #define KLSK_INLINE_H KLSK_INLINE inline
 #define KLSK_INLINE_CE KLSK_INLINE constexpr
 
-// TODO: using `#pragma GCC unroll`
+#define KLSK_STRING(x) #x
+
+#if defined(__clang__)
+  #define KLSK_UNROLL(N) _Pragma(KLSK_STRING(unroll N))
+#elif defined(__GNUC__)
+  #define KLSK_UNROLL(N) _Pragma(KLSK_STRING(GCC unroll N))
+#else
+  #define KLSK_UNROLL(N)
+#endif
 
 /// Prevent reordering for both compiler and processor.
 #define KLSK_MEM_BARRIER std::atomic_thread_fence(std::memory_order_seq_cst)