optimize compile time

CMakeLists.txt

@@ -0,0 +1,29 @@
+cmake_minimum_required(VERSION 3.14)
+project(wjr 
+    LANGUAGES CXX
+)
+
+if(NOT CMAKE_BUILD_TYPE)
+   set(CMAKE_BUILD_TYPE Release)
+endif()
+
+set(CMAKE_CXX_FLAGS "$ENV{CXXFLAGS} -std=c++17 -march=native -Wall -Wextra")
+
+set(WJR_INCLUDE_DIR ${PROJECT_SOURCE_DIR}/include)
+set(WJR_LIB_DIR ${PROJECT_SOURCE_DIR}/src)
+
+file(GLOB_RECURSE WJR_SRCS ${WJR_LIB_DIR}/*.cpp)
+
+target_include_directories(wjr PUBLIC
+   ${WJR_INCLUDE_DIR}
+)
+
+target_precompile_headers(
+   wjr
+   PUBLIC ${WJR_INCLUDE_DIR}/wjr/preprocessor.hpp
+   PUBLIC ${WJR_INCLUDE_DIR}/wjr/assert.hpp
+   PUBLIC ${WJR_INCLUDE_DIR}/wjr/type_traits.hpp
+   PUBLIC ${WJR_INCLUDE_DIR}/wjr/x86/simd/simd.hpp
+)
+
+add_library(wjr STATIC ${WJR_SRCS})

include/wjr/assert.hpp

@@ -27,17 +27,10 @@
  *
  */
 
-#include <cstdio>
-#include <cstdlib>
 #include <iostream>
-#include <tuple>
-#include <type_traits>
-#include <utility>
 
 #include <wjr/preprocessor.hpp>
 
-namespace wjr {
-
 #ifndef WJR_DEBUG_LEVEL
 #if defined(NDEBUG)
 #define WJR_DEBUG_LEVEL 0
@@ -50,6 +43,8 @@ namespace wjr {
 #error "WJR_DEBUG_LEVEL must be 0 ~ 3"
 #endif
 
+namespace wjr {
+
 #define WJR_DEBUG_IF(level, expr0, expr1)                                                \
     WJR_PP_BOOL_IF(WJR_PP_GT(WJR_DEBUG_LEVEL, level), expr0, expr1)
 

include/wjr/biginteger/biginteger.hpp

@@ -134,21 +134,21 @@ class default_biginteger_vector_storage {
     ~default_biginteger_vector_storage() noexcept = default;
 
     void destroy(_Alty &al) noexcept {
-        if (WJR_BUILTIN_CONSTANT_P(data() == nullptr) && data() == nullptr) {
+        if (WJR_BUILTIN_CONSTANT_P_TRUE(data() == nullptr)) {
             return;
         }
 
         const size_type __size = size();
 
-        if (WJR_BUILTIN_CONSTANT_P(__size == 0) && __size == 0) {
+        if (WJR_BUILTIN_CONSTANT_P_TRUE(__size == 0)) {
             return;
         }
 
         destroy_n_using_allocator(data(), __size, al);
     }
 
     void destroy_and_deallocate(_Alty &al) noexcept {
-        if (WJR_BUILTIN_CONSTANT_P(capacity() == 0) && capacity() == 0) {
+        if (WJR_BUILTIN_CONSTANT_P_TRUE(capacity() == 0)) {
             return;
         }
 
@@ -237,15 +237,15 @@ inline int32_t __compare_si_impl(const biginteger_data *lhs, int64_t rhs) noexce
 /// @private
 template <typename T, WJR_REQUIRES(is_nonbool_integral_v<T>)>
 int32_t __compare_impl(const biginteger_data *lhs, T rhs) noexcept {
-    if (WJR_BUILTIN_CONSTANT_P(rhs == 0) && rhs == 0) {
+    if (WJR_BUILTIN_CONSTANT_P_TRUE(rhs == 0)) {
         const int32_t ssize = lhs->get_ssize();
         return ssize == 0 ? 0 : ssize < 0 ? -1 : 1;
     }
 
     if constexpr (std::is_unsigned_v<T>) {
         return __compare_ui_impl(lhs, rhs);
     } else {
-        if (WJR_BUILTIN_CONSTANT_P(rhs >= 0) && rhs >= 0) {
+        if (WJR_BUILTIN_CONSTANT_P_TRUE(rhs >= 0)) {
             return __compare_ui_impl(lhs, to_unsigned(rhs));
         }
 
@@ -770,8 +770,7 @@ void sqr(basic_biginteger<S> &dst, const biginteger_data &src) noexcept {
 template <typename S>
 void mul(basic_biginteger<S> &dst, const biginteger_data &lhs,
          const biginteger_data &rhs) noexcept {
-    if (WJR_BUILTIN_CONSTANT_P(std::addressof(lhs) == std::addressof(rhs)) &&
-        std::addressof(lhs) == std::addressof(rhs)) {
+    if (WJR_BUILTIN_CONSTANT_P_TRUE(std::addressof(lhs) == std::addressof(rhs))) {
         sqr(dst, lhs);
         return;
     }

include/wjr/container/generic/details.hpp

@@ -7,22 +7,6 @@
 
 namespace wjr {
 
-WJR_REGISTER_HAS_TYPE(container_begin,
-                      std::begin(std::declval<std::add_lvalue_reference_t<Container>>()),
-                      Container);
-WJR_REGISTER_HAS_TYPE(container_cbegin,
-                      std::cbegin(std::declval<std::add_lvalue_reference_t<Container>>()),
-                      Container);
-WJR_REGISTER_HAS_TYPE(container_end,
-                      std::end(std::declval<std::add_lvalue_reference_t<Container>>()),
-                      Container);
-WJR_REGISTER_HAS_TYPE(container_cend,
-                      std::cend(std::declval<std::add_lvalue_reference_t<Container>>()),
-                      Container);
-WJR_REGISTER_HAS_TYPE(container_size,
-                      std::size(std::declval<std::add_lvalue_reference_t<Container>>()),
-                      Container);
-
 WJR_REGISTER_HAS_TYPE(__container_resize,
                       std::declval<std::add_lvalue_reference_t<Container>>().resize(
                           std::declval<Size>(), std::declval<Args>()...),

include/wjr/container/generic/vector.hpp

@@ -282,11 +282,11 @@ class __default_vector_storage_impl {
 
     WJR_CONSTEXPR20 void
     destroy(_Alty &al) noexcept(std::is_nothrow_destructible_v<value_type>) {
-        if (WJR_BUILTIN_CONSTANT_P(data() == nullptr) && data() == nullptr) {
+        if (WJR_BUILTIN_CONSTANT_P_TRUE(data() == nullptr)) {
             return;
         }
 
-        if (WJR_BUILTIN_CONSTANT_P(size() == 0) && size() == 0) {
+        if (WJR_BUILTIN_CONSTANT_P_TRUE(size() == 0)) {
             return;
         }
 
@@ -295,11 +295,11 @@ class __default_vector_storage_impl {
 
     WJR_CONSTEXPR20 void destroy_and_deallocate(_Alty &al) noexcept(
         std::is_nothrow_destructible_v<value_type>) {
-        if (WJR_BUILTIN_CONSTANT_P(data() == nullptr) && data() == nullptr) {
+        if (WJR_BUILTIN_CONSTANT_P_TRUE(data() == nullptr)) {
             return;
         }
 
-        if (WJR_BUILTIN_CONSTANT_P(capacity() == 0) && capacity() == 0) {
+        if (WJR_BUILTIN_CONSTANT_P_TRUE(capacity() == 0)) {
             return;
         }
 
@@ -404,7 +404,7 @@ class __static_vector_storage_impl {
 
     WJR_CONSTEXPR20 void
     destroy(_Alty &al) noexcept(std::is_nothrow_destructible_v<value_type>) {
-        if (WJR_BUILTIN_CONSTANT_P(size() == 0) && size() == 0) {
+        if (WJR_BUILTIN_CONSTANT_P_TRUE(size() == 0)) {
             return;
         }
 
@@ -558,15 +558,15 @@ class __fixed_vector_storage_impl {
 
 private:
     WJR_PURE WJR_INTRINSIC_INLINE bool __is_null_data() const {
-        return WJR_BUILTIN_CONSTANT_P(data() == nullptr) && data() == nullptr;
+        return WJR_BUILTIN_CONSTANT_P_TRUE(data() == nullptr);
     }
 
     WJR_PURE WJR_INTRINSIC_INLINE bool __is_zero_size() const {
-        return WJR_BUILTIN_CONSTANT_P(size() == 0) && size() == 0;
+        return WJR_BUILTIN_CONSTANT_P_TRUE(size() == 0);
     }
 
     WJR_PURE WJR_INTRINSIC_INLINE bool __is_zero_capacity() const {
-        return WJR_BUILTIN_CONSTANT_P(capacity() == 0) && capacity() == 0;
+        return WJR_BUILTIN_CONSTANT_P_TRUE(capacity() == 0);
     }
 
 public:
@@ -712,7 +712,7 @@ class __sso_vector_storage_impl {
 
     WJR_CONSTEXPR20 void
     destroy(_Alty &al) noexcept(std::is_nothrow_destructible_v<value_type>) {
-        if (WJR_BUILTIN_CONSTANT_P(size() == 0) && size() == 0) {
+        if (WJR_BUILTIN_CONSTANT_P_TRUE(size() == 0)) {
             return;
         }
 
@@ -1379,8 +1379,8 @@ class basic_vector {
     template <typename Iter, WJR_REQUIRES(is_iterator_v<Iter>)>
     WJR_CONSTEXPR20 iterator insert(const_iterator pos, Iter first, Iter last) {
         const auto old_pos = static_cast<size_type>(pos - cbegin());
-        __range_insert(data() + old_pos, to_contiguous_address(first), to_contiguous_address(last),
-                       iterator_category_t<Iter>());
+        __range_insert(data() + old_pos, to_contiguous_address(first),
+                       to_contiguous_address(last), iterator_category_t<Iter>());
         return begin() + old_pos;
     }
 
@@ -1500,8 +1500,9 @@ class basic_vector {
     template <typename Iter, WJR_REQUIRES(is_iterator_v<Iter>)>
     WJR_CONSTEXPR20 basic_vector &replace(const_iterator from, const_iterator to,
                                           Iter first, Iter last) {
-        __range_replace(__get_pointer(from), __get_pointer(to), to_contiguous_address(first),
-                        to_contiguous_address(last), iterator_category_t<Iter>());
+        __range_replace(__get_pointer(from), __get_pointer(to),
+                        to_contiguous_address(first), to_contiguous_address(last),
+                        iterator_category_t<Iter>());
         return *this;
     }
 

include/wjr/generic/math/bignum-config.hpp

@@ -0,0 +1,72 @@
+#ifndef WJR_GENERIC_MATH_BIGNUM_CONFIG_HPP__
+#define WJR_GENERIC_MATH_BIGNUM_CONFIG_HPP__
+
+#ifndef WJR_TOOM22_MUL_THRESHOLD
+#define WJR_TOOM22_MUL_THRESHOLD 22
+#endif
+
+#ifndef WJR_TOOM33_MUL_THRESHOLD
+#define WJR_TOOM33_MUL_THRESHOLD 84
+#endif
+
+#ifndef WJR_TOOM44_MUL_THRESHOLD
+#define WJR_TOOM44_MUL_THRESHOLD 208
+#endif
+
+#ifndef WJR_TOOM55_MUL_THRESHOLD
+#define WJR_TOOM55_MUL_THRESHOLD 800
+#endif
+
+#ifndef WJR_TOOM32_TO_TOOM43_MUL_THRESHOLD
+#define WJR_TOOM32_TO_TOOM43_MUL_THRESHOLD 73
+#endif
+
+#ifndef WJR_TOOM32_TO_TOOM53_MUL_THRESHOLD
+#define WJR_TOOM32_TO_TOOM53_MUL_THRESHOLD 153
+#endif
+
+#ifndef WJR_TOOM42_TO_TOOM53_MUL_THRESHOLD
+#define WJR_TOOM42_TO_TOOM53_MUL_THRESHOLD 137
+#endif
+
+#ifndef WJR_TOOM42_TO_TOOM63_MUL_THRESHOLD
+#define WJR_TOOM42_TO_TOOM63_MUL_THRESHOLD 153
+#endif
+
+#ifndef WJR_TOOM2_SQR_THRESHOLD
+#define WJR_TOOM2_SQR_THRESHOLD 34
+#endif
+
+#ifndef WJR_TOOM3_SQR_THRESHOLD
+#define WJR_TOOM3_SQR_THRESHOLD 124
+#endif
+
+#ifndef WJR_TOOM4_SQR_THRESHOLD
+#define WJR_TOOM4_SQR_THRESHOLD 288
+#endif
+
+#ifndef WJR_TOOM5_SQR_THRESHOLD
+#define WJR_TOOM5_SQR_THRESHOLD 980
+#endif
+
+#ifndef WJR_DC_DIV_QR_THRESHOLD
+#define WJR_DC_DIV_QR_THRESHOLD (WJR_TOOM22_MUL_THRESHOLD * 2)
+#endif // WJR_DC_DIV_QR_THRESHOLD
+
+#ifndef WJR_DC_BIGNUM_TO_CHARS_THRESHOLD
+#define WJR_DC_BIGNUM_TO_CHARS_THRESHOLD 20
+#endif
+
+#ifndef WJR_DC_BIGNUM_TO_CHARS_PRECOMPUTE_THRESHOLD
+#define WJR_DC_BIGNUM_TO_CHARS_PRECOMPUTE_THRESHOLD 20
+#endif
+
+#ifndef WJR_DC_BIGNUM_FROM_CHARS_THRESHOLD
+#define WJR_DC_BIGNUM_FROM_CHARS_THRESHOLD 1670
+#endif
+
+#ifndef WJR_DC_BIGNUM_FROM_CHARS_PRECOMPUTE_THRESHOLD
+#define WJR_DC_BIGNUM_FROM_CHARS_PRECOMPUTE_THRESHOLD 3105
+#endif
+
+#endif // WJR_GENERIC_MATH_BIGNUM_CONFIG_HPP__

include/wjr/math/add.hpp

@@ -214,7 +214,7 @@ WJR_INTRINSIC_CONSTEXPR_E U addc_1(uint64_t *dst, const uint64_t *src0, size_t n
     WJR_ASSERT_L2(WJR_IS_SAME_OR_INCR_P(dst, n, src0, n));
     WJR_ASSERT_ASSUME(c_in <= 1);
 
-    if (WJR_BUILTIN_CONSTANT_P(n == 1) && n == 1) {
+    if (WJR_BUILTIN_CONSTANT_P_TRUE(n == 1)) {
         uint8_t overflow = 0;
         dst[0] = addc_cc(src0[0], src1, c_in, overflow);
         return static_cast<U>(overflow);
@@ -377,8 +377,8 @@ WJR_INTRINSIC_CONSTEXPR_E void __add_128(uint64_t &al, uint64_t &ah, uint64_t lo
                                          uint64_t hi0, uint64_t lo1,
                                          uint64_t hi1) noexcept {
 #if WJR_HAS_BUILTIN(__BUILTIN_ADD_128) || WJR_HAS_BUILTIN(__ASM_ADD_128)
-    if (is_constant_evaluated() || (WJR_BUILTIN_CONSTANT_P(lo0 == 0) && lo0 == 0) ||
-        (WJR_BUILTIN_CONSTANT_P(lo1 == 0) && lo1 == 0)) {
+    if (is_constant_evaluated() || WJR_BUILTIN_CONSTANT_P_TRUE(lo0 == 0) ||
+        WJR_BUILTIN_CONSTANT_P_TRUE(lo1 == 0) || WJR_BUILTIN_CONSTANT_P(lo0 + lo1)) {
         return __fallback_add_128(al, ah, lo0, hi0, lo1, hi1);
     }
 

include/wjr/math/bignum-config.hpp

@@ -1,72 +1,6 @@
 #ifndef WJR_MATH_BIGNUM_CONFIG_HPP__
 #define WJR_MATH_BIGNUM_CONFIG_HPP__
 
-#ifndef WJR_TOOM22_MUL_THRESHOLD
-#define WJR_TOOM22_MUL_THRESHOLD 22
-#endif
-
-#ifndef WJR_TOOM33_MUL_THRESHOLD
-#define WJR_TOOM33_MUL_THRESHOLD 84
-#endif
-
-#ifndef WJR_TOOM44_MUL_THRESHOLD
-#define WJR_TOOM44_MUL_THRESHOLD 208
-#endif
-
-#ifndef WJR_TOOM55_MUL_THRESHOLD
-#define WJR_TOOM55_MUL_THRESHOLD 800
-#endif
-
-#ifndef WJR_TOOM32_TO_TOOM43_MUL_THRESHOLD
-#define WJR_TOOM32_TO_TOOM43_MUL_THRESHOLD 73
-#endif
-
-#ifndef WJR_TOOM32_TO_TOOM53_MUL_THRESHOLD
-#define WJR_TOOM32_TO_TOOM53_MUL_THRESHOLD 153
-#endif
-
-#ifndef WJR_TOOM42_TO_TOOM53_MUL_THRESHOLD
-#define WJR_TOOM42_TO_TOOM53_MUL_THRESHOLD 137
-#endif
-
-#ifndef WJR_TOOM42_TO_TOOM63_MUL_THRESHOLD
-#define WJR_TOOM42_TO_TOOM63_MUL_THRESHOLD 153
-#endif
-
-#ifndef WJR_TOOM2_SQR_THRESHOLD
-#define WJR_TOOM2_SQR_THRESHOLD 34
-#endif
-
-#ifndef WJR_TOOM3_SQR_THRESHOLD
-#define WJR_TOOM3_SQR_THRESHOLD 124
-#endif
-
-#ifndef WJR_TOOM4_SQR_THRESHOLD
-#define WJR_TOOM4_SQR_THRESHOLD 288
-#endif
-
-#ifndef WJR_TOOM5_SQR_THRESHOLD
-#define WJR_TOOM5_SQR_THRESHOLD 980
-#endif
-
-#ifndef WJR_DC_DIV_QR_THRESHOLD
-#define WJR_DC_DIV_QR_THRESHOLD (WJR_TOOM22_MUL_THRESHOLD * 2)
-#endif // WJR_DC_DIV_QR_THRESHOLD
-
-#ifndef WJR_DC_BIGNUM_TO_CHARS_THRESHOLD
-#define WJR_DC_BIGNUM_TO_CHARS_THRESHOLD 20
-#endif
-
-#ifndef WJR_DC_BIGNUM_TO_CHARS_PRECOMPUTE_THRESHOLD
-#define WJR_DC_BIGNUM_TO_CHARS_PRECOMPUTE_THRESHOLD 20
-#endif
-
-#ifndef WJR_DC_BIGNUM_FROM_CHARS_THRESHOLD
-#define WJR_DC_BIGNUM_FROM_CHARS_THRESHOLD 1670
-#endif
-
-#ifndef WJR_DC_BIGNUM_FROM_CHARS_PRECOMPUTE_THRESHOLD
-#define WJR_DC_BIGNUM_FROM_CHARS_PRECOMPUTE_THRESHOLD 3105
-#endif
+#include <wjr/generic/math/bignum-config.hpp>
 
 #endif // WJR_MATH_BIGNUM_CONFIG_HPP__