simdjson · jkeiser · Aug 10, 2024
diff --git a/.clangd b/.clangd
@@ -45,3 +45,11 @@ Diagnostics:
     Suppress:
         - pragma_attribute_no_pop_eof
         - pragma_attribute_stack_mismatch
+---
+# clang 18
+If:
+    PathMatch:
+        - dependencies/.cache/json11/json11.cpp
+CompileFlags:
+    Add:
+        - -Wno-unqualified-std-cast-call
diff --git a/include/simdjson/arm64/simd.h b/include/simdjson/arm64/simd.h
@@ -412,6 +412,66 @@ namespace {
     }
   };
 
+#ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
+      static const uint8x16_t BITMASK64_BUILDER_MASK = simdjson_make_uint8x16_t(
+        0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
+        0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80
+      );
+#else
+      static const uint8x16_t BITMASK64_BUILDER_MASK = {
+        0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
+        0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80
+      };
+#endif
+
+  template <int N = 0>
+  struct simd_bitmask64_builder;
+
+  template<>
+  struct simd_bitmask64_builder<4> {
+    const uint64_t mask;
+    operator uint64_t() && { return mask; }
+  }; // struct simd_bitmask64_builder<4>
+
+  template<>
+  struct simd_bitmask64_builder<3> {
+    const uint8x16_t sum01;
+    const simd8<bool> val2;
+    simdjson_inline simd_bitmask64_builder<4> next(simd8<bool> val3) {
+      // Add each of the elements next to each other, successively, to stuff each 8 byte mask into one.
+      uint8x16_t sum23 = vpaddq_u8(val2 & BITMASK64_BUILDER_MASK, val3 & BITMASK64_BUILDER_MASK);
+      uint8x16_t sum0123 = vpaddq_u8(sum01, sum23);
+
+      // This algorithm is actually designed to create a 128-bit mask from 8 16-byte simd masks,
+      // but since we only want 64 bits, we add the mask to itself (creating the final mask twice).
+      uint8x16_t sum01230123 = vpaddq_u8(sum0123, sum0123);
+      return { vgetq_lane_u64(vreinterpretq_u64_u8(sum01230123), 0) };
+    }
+  }; // struct simd_bitmask64_builder<3>
+
+  template<>
+  struct simd_bitmask64_builder<2> {
+    const uint8x16_t sum01;
+    simdjson_inline simd_bitmask64_builder<3> next(simd8<bool> val2) {
+      return { sum01, val2 };
+    }
+  }; // struct simd_bitmask64_builder<2>
+
+  template<>
+  struct simd_bitmask64_builder<1> {
+    const simd8<bool> val0;
+    simdjson_inline simd_bitmask64_builder<2> next(simd8<bool> val1) {
+      return { vpaddq_u8(val0 & BITMASK64_BUILDER_MASK, val1 & BITMASK64_BUILDER_MASK) };
+    }
+  }; // struct simd_bitmask64_builder<1>
+
+  template<>
+  struct simd_bitmask64_builder<0> {
+    simdjson_inline simd_bitmask64_builder<1> next(simd8<bool> val) {
+      return { val };
+    }
+  }; // struct simd_bitmask64_builder<0>
+
   template<typename T>
   struct simd8x64 {
     static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
@@ -449,23 +509,11 @@ namespace {
     }
 
     simdjson_inline uint64_t to_bitmask() const {
-#ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
-      const uint8x16_t bit_mask = simdjson_make_uint8x16_t(
-        0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
-        0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80
-      );
-#else
-      const uint8x16_t bit_mask = {
-        0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
-        0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80
-      };
-#endif
-      // Add each of the elements next to each other, successively, to stuff each 8 byte mask into one.
-      uint8x16_t sum0 = vpaddq_u8(this->chunks[0] & bit_mask, this->chunks[1] & bit_mask);
-      uint8x16_t sum1 = vpaddq_u8(this->chunks[2] & bit_mask, this->chunks[3] & bit_mask);
-      sum0 = vpaddq_u8(sum0, sum1);
-      sum0 = vpaddq_u8(sum0, sum0);
-      return vgetq_lane_u64(vreinterpretq_u64_u8(sum0), 0);
+      return simd_bitmask64_builder<0>()
+        .next(this->chunks[0])
+        .next(this->chunks[1])
+        .next(this->chunks[2])
+        .next(this->chunks[3]);
     }
 
     simdjson_inline uint64_t eq(const T m) const {

diff --git a/include/simdjson/haswell/simd.h b/include/simdjson/haswell/simd.h
@@ -295,6 +295,30 @@ namespace simd {
     simdjson_inline int get_bit() const { return _mm256_movemask_epi8(_mm256_slli_epi16(*this, 7-N)); }
   };
 
+  template <int N = 0>
+  struct simd_bitmask64_builder;
+
+  template<>
+  struct simd_bitmask64_builder<2> {
+    const uint64_t bitmask;
+    operator uint64_t() && { return bitmask; }
+  }; // struct simd_bitmask64_builder<2>
+
+  template<>
+  struct simd_bitmask64_builder<1> {
+    const int bitmask;
+    simdjson_inline simd_bitmask64_builder<2> next(simd8<bool> val) {
+      uint64_t r_lo = uint32_t(bitmask);
+      uint64_t r_hi = val.to_bitmask();
+      return { r_lo | (r_hi << 32) };
+    }
+  }; // struct simd_bitmask64_builder<1>
+
+  template<>
+  struct simd_bitmask64_builder<0> {
+    simdjson_inline simd_bitmask64_builder<1> next(simd8<bool> val) { return { val.to_bitmask() }; }
+  }; // struct simd_bitmask64_builder<0>
+
   template<typename T>
   struct simd8x64 {
     static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
@@ -322,9 +346,9 @@ namespace simd {
     }
 
     simdjson_inline uint64_t to_bitmask() const {
-      uint64_t r_lo = uint32_t(this->chunks[0].to_bitmask());
-      uint64_t r_hi =                       this->chunks[1].to_bitmask();
-      return r_lo | (r_hi << 32);
+      return simd_bitmask64_builder<0>()
+        .next(this->chunks[0])
+        .next(this->chunks[1]);
     }
 
     simdjson_inline simd8<T> reduce_or() const {

diff --git a/include/simdjson/icelake/simd.h b/include/simdjson/icelake/simd.h
@@ -314,6 +314,20 @@ namespace simd {
     simdjson_inline uint64_t get_bit() const { return _mm512_movepi8_mask(_mm512_slli_epi16(*this, 7-N)); }
   };
 
+  template <int N = 0>
+  struct simd_bitmask64_builder;
+
+  template<>
+  struct simd_bitmask64_builder<1> {
+    const __mmask64 bitmask;
+    operator __mmask64() && { return bitmask; }
+  }; // struct simd_bitmask64_builder<2>
+
+  template<>
+  struct simd_bitmask64_builder<0> {
+    simdjson_inline simd_bitmask64_builder<1> next(simd8<bool> val) { return { _mm512_movepi8_mask(val) }; }
+  }; // struct simd_bitmask64_builder<0>
+
   template<typename T>
   struct simd8x64 {
     static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);

diff --git a/include/simdjson/lasx/simd.h b/include/simdjson/lasx/simd.h
@@ -298,6 +298,35 @@ namespace simd {
     simdjson_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(__lasx_xvslli_b(*this, N)); }
   };
 
+  template <int N = 0>
+  struct simd_bitmask64_builder;
+
+  template<>
+  struct simd_bitmask64_builder<2> {
+    const unsigned long int mask01;
+    operator uint64_t() { return mask01; }
+  }; // struct simd_bitmask64_builder<2>
+
+  template<>
+  struct simd_bitmask64_builder<1> {
+    const __m256i mask0;
+    simdjson_inline simd_bitmask64_builder<2> next(simd8<bool> val1) {
+      __m256i mask1 = __lasx_xvmskltz_b(val1);
+      __m256i mask_tmp = __lasx_xvpickve_w(mask0, 4);
+      __m256i tmp = __lasx_xvpickve_w(mask1, 4);
+      __m256i mask01 = __lasx_xvinsve0_w(mask0, mask1, 1);
+      __m256i mask01_tmp = __lasx_xvinsve0_w(mask_tmp, tmp, 1);
+      return { __lasx_xvpickve2gr_du(__lasx_xvpackev_h(mask01_tmp, mask01), 0) };
+    }
+  }; // struct simd_bitmask64_builder<1>
+
+  template<>
+  struct simd_bitmask64_builder<0> {
+    simdjson_inline simd_bitmask64_builder<1> next(simd8<bool> val) {
+      return { __lasx_xvmskltz_b(val0) };
+    }
+  }; // struct simd_bitmask64_builder<0>
+
   template<typename T>
   struct simd8x64 {
     static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
@@ -331,13 +360,9 @@ namespace simd {
     }
 
     simdjson_inline uint64_t to_bitmask() const {
-      __m256i mask0 = __lasx_xvmskltz_b(this->chunks[0]);
-      __m256i mask1 = __lasx_xvmskltz_b(this->chunks[1]);
-      __m256i mask_tmp = __lasx_xvpickve_w(mask0, 4);
-      __m256i tmp = __lasx_xvpickve_w(mask1, 4);
-      mask0 = __lasx_xvinsve0_w(mask0, mask1, 1);
-      mask_tmp = __lasx_xvinsve0_w(mask_tmp, tmp, 1);
-      return __lasx_xvpickve2gr_du(__lasx_xvpackev_h(mask_tmp, mask0), 0);
+      return simd_bitmask64_builder<0>()
+        .next(this->chunks[0])
+        .next(this->chunks[1]);
     }
 
     simdjson_inline simd8<T> reduce_or() const {

diff --git a/include/simdjson/lsx/simd.h b/include/simdjson/lsx/simd.h
@@ -255,6 +255,50 @@ namespace simd {
     simdjson_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(__lsx_vslli_b(*this, N)); }
   };
 
+  template <int N = 0>
+  struct simd_bitmask64_builder;
+
+  template<>
+  struct simd_bitmask64_builder<4> {
+    const unsigned long int result;
+    operator uint64_t() { return result; }
+  }; // struct simd_bitmask64_builder<4>
+
+  template<>
+  struct simd_bitmask64_builder<3> {
+    const __m128i mask01;
+    const __m128i mask2;
+    simdjson_inline simd_bitmask64_builder<4> next(simd8<bool> val3) {
+      __m128i mask3 = __lsx_vmskltz_b(val3);
+      __m128i mask23 = __lsx_vilvl_h(mask3, mask2);
+      return { __lsx_vpickve2gr_du(__lsx_vilvl_w(mask23, mask01), 0) };
+    }
+  }; // struct simd_bitmask64_builder<3>
+
+  template<>
+  struct simd_bitmask64_builder<2> {
+    const __m128i mask01;
+    simdjson_inline simd_bitmask64_builder<3> next(simd8<bool> val2) {
+      return { mask01, __lsx_vmskltz_b(val2) };
+    }
+  }; // struct simd_bitmask64_builder<2>
+
+  template<>
+  struct simd_bitmask64_builder<1> {
+    const __m128i mask0;
+    simdjson_inline simd_bitmask64_builder<2> next(simd8<bool> val1) {
+      __m128i mask1 = __lsx_vmskltz_b(val1);
+      return { __lsx_vilvl_h(mask1, mask0) };
+    }
+  }; // struct simd_bitmask64_builder<1>
+
+  template<>
+  struct simd_bitmask64_builder<0> {
+    simdjson_inline simd_bitmask64_builder<1> next(simd8<bool> val0) {
+      return { __lsx_vmskltz_b(val0) };
+    }
+  }; // struct simd_bitmask64_builder<0>
+
   template<typename T>
   struct simd8x64 {
     static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
@@ -303,13 +347,11 @@ namespace simd {
     }
 
     simdjson_inline uint64_t to_bitmask() const {
-      __m128i mask1 = __lsx_vmskltz_b(this->chunks[0]);
-      __m128i mask2 = __lsx_vmskltz_b(this->chunks[1]);
-      __m128i mask3 = __lsx_vmskltz_b(this->chunks[2]);
-      __m128i mask4 = __lsx_vmskltz_b(this->chunks[3]);
-      mask1 = __lsx_vilvl_h(mask2, mask1);
-      mask2 = __lsx_vilvl_h(mask4, mask3);
-      return __lsx_vpickve2gr_du(__lsx_vilvl_w(mask2, mask1), 0);
+      return simd_bitmask64_builder<0>()
+        .next(this->chunks[0])
+        .next(this->chunks[1])
+        .next(this->chunks[2])
+        .next(this->chunks[3]);
     }
 
     simdjson_inline simd8<T> reduce_or() const {

diff --git a/include/simdjson/ppc64/simd.h b/include/simdjson/ppc64/simd.h
@@ -392,6 +392,51 @@ template <> struct simd8<uint8_t> : base8_numeric<uint8_t> {
   }
 };
 
+template <int N = 0>
+struct simd_bitmask64_builder;
+
+template<>
+struct simd_bitmask64_builder<4> {
+  const uint64_t bitmask;
+  simdjson_inline operator uint64_t() && { return bitmask; }
+}; // struct simd_bitmask64_builder<4>
+
+template<>
+struct simd_bitmask64_builder<3> {
+  const int bitmask01;
+  const int bitmask2;
+  simdjson_inline simd_bitmask64_builder<4> next(simd8<bool> val) {
+    return {
+      uint64_t(this->bitmask01) |
+      (uint64_t(this->bitmask2) << 32) |
+      (uint64_t(val.to_bitmask()) << 48)
+    };
+  }
+}; // struct simd_bitmask64_builder<3>
+
+template<>
+struct simd_bitmask64_builder<2> {
+  const int bitmask01;
+  simdjson_inline simd_bitmask64_builder<3> next(simd8<bool> val) {
+    return { bitmask01, val.to_bitmask() };
+  }
+}; // struct simd_bitmask64_builder<2>
+
+template<>
+struct simd_bitmask64_builder<1> {
+  const int bitmask0;
+  simdjson_inline simd_bitmask64_builder<2> next(simd8<bool> val) {
+    return { bitmask0 | (val.to_bitmask() << 16) };
+  }
+}; // struct simd_bitmask64_builder<1>
+
+template<>
+struct simd_bitmask64_builder<0> {
+  simdjson_inline simd_bitmask64_builder<1> next(simd8<bool> val) {
+    return { val.to_bitmask() };
+  }
+}; // struct simd_bitmask64_builder<0>
+
 template <typename T> struct simd8x64 {
   static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
   static_assert(NUM_CHUNKS == 4,
@@ -434,11 +479,11 @@ template <typename T> struct simd8x64 {
   }
 
   simdjson_inline uint64_t to_bitmask() const {
-    uint64_t r0 = uint32_t(this->chunks[0].to_bitmask());
-    uint64_t r1 = this->chunks[1].to_bitmask();
-    uint64_t r2 = this->chunks[2].to_bitmask();
-    uint64_t r3 = this->chunks[3].to_bitmask();
-    return r0 | (r1 << 16) | (r2 << 32) | (r3 << 48);
+      return simd_bitmask64_builder<0>()
+        .next(this->chunks[0])
+        .next(this->chunks[1])
+        .next(this->chunks[2])
+        .next(this->chunks[3]);
   }
 
   simdjson_inline uint64_t eq(const T m) const {