jp/horizontal sum

inc/zoo/meta/popcount.h

@@ -2,6 +2,8 @@
 #define ZOO_HEADER_META_POPCOUNT_H
 
 #include "zoo/meta/BitmaskMaker.h"
+#include <cstdint>
+#include <type_traits>
 
 namespace zoo { namespace meta {
 
@@ -90,6 +92,30 @@ struct PopcountIntrinsic {
 };
 #endif
 
+template<typename T>
+constexpr auto NumBits() {
+    return sizeof(T) * 8;
+}
+static_assert(NumBits<uint16_t>() == 16);
+static_assert(NumBits<uint64_t>() == 64);
+
+template <typename T = std::uint64_t>
+constexpr
+std::enable_if_t<std::is_unsigned_v<T> && NumBits<T>() <= 64, T>
+basic_popcount(T x) {
+    if constexpr (NumBits<T>() <= 32) {
+        return __builtin_popcountl(x);
+    } else {
+        return __builtin_popcountll(x);
+    }
+}
+
+static_assert(basic_popcount<uint64_t>(0b111) == 3);
+static_assert(basic_popcount<uint64_t>(0xFF) == 8);
+static_assert(basic_popcount<uint64_t>(0xFF'FF'FF'FF) == 32);
+static_assert(basic_popcount<uint64_t>(0xFF'FF'FF'FF'FF'FF'FF'FF) == 64);
+static_assert(basic_popcount<uint64_t>(0xFF'FF'FF'FF'FF'FF'FF'FF - 2 - 4 - 8) == 61);
+
 }}
 
 #endif

inc/zoo/swar/SWAR.h

@@ -299,7 +299,7 @@ constexpr auto broadcast(SWAR<NBits, T> v) {
 
 /// BooleanSWAR treats the MSB of each SWAR lane as the boolean associated with that lane.
 template<int NBits, typename T>
-struct BooleanSWAR: SWAR<NBits, T> {
+struct BooleanSWAR : SWAR<NBits, T> {
     using Base = SWAR<NBits, T>;
 
     template<std::size_t N, typename = std::enable_if_t<Base::Lanes == N>>
@@ -308,8 +308,13 @@ struct BooleanSWAR: SWAR<NBits, T> {
     { this->m_v <<= (NBits - 1); }
 
     // Booleanness is stored in the MSBs
-    static constexpr auto MaskMSB =
-        broadcast<NBits, T>(Base(T(1) << (NBits -1)));
+    static constexpr auto MaskMSB = []{
+        if constexpr (SWAR<NBits, T>::Lanes == 1) {
+            return Base(T{~0}); // all on, no lanes
+        }
+        return broadcast<NBits, T>(Base(T(1) << (NBits - 1)));
+    }();
+
     static constexpr auto AllTrue = MaskMSB;
     static constexpr auto MaskLSB =
          broadcast<NBits, T>(Base(T(1)));
@@ -392,6 +397,11 @@ BooleanSWAR(
     const bool (&values)[BooleanSWAR<NBits, T>::Lanes]
 ) -> BooleanSWAR<NBits, T>;
 
+template <int NBits, typename T>
+BooleanSWAR(
+    SWAR<NBits, T> arg
+) -> BooleanSWAR<NBits, T>;
+
 template<int NBits, typename T>
 constexpr BooleanSWAR<NBits, T>
 convertToBooleanSWAR(SWAR<NBits, T> arg) noexcept {
@@ -586,6 +596,10 @@ constexpr SWAR<NBits, T> logarithmFloor(SWAR<NBits, T> v) noexcept {
     return SWAR<NBits, T>{popcounts - ones};
 }
 
+
+
+
+
 static_assert(
     logarithmFloor(SWAR<8>{0x8040201008040201ull}).value() ==
     0x0706050403020100ull