/*
The MIT License (MIT)

Copyright (c) 2012-Present, Syoyo Fujita and many contributors.

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.
*/

//
// version 2.0.0 : Add new object oriented API. 1.x API is still provided.
//                 * Add python binding.
//                 * Support line primitive.
//                 * Support points primitive.
//                 * Support multiple search path for .mtl(v1 API).
//                 * Support vertex skinning weight `vw`(as an tinyobj
//                 extension). Note that this differs vertex weight([w]
//                 component in `v` line)
//                 * Support escaped whitespece in mtllib
//                 * Add robust triangulation using Mapbox
//                 earcut(TINYOBJLOADER_USE_MAPBOX_EARCUT).
// version 1.4.0 : Modifed ParseTextureNameAndOption API
// version 1.3.1 : Make ParseTextureNameAndOption API public
// version 1.3.0 : Separate warning and error message(breaking API of LoadObj)
// version 1.2.3 : Added color space extension('-colorspace') to tex opts.
// version 1.2.2 : Parse multiple group names.
// version 1.2.1 : Added initial support for line('l') primitive(PR #178)
// version 1.2.0 : Hardened implementation(#175)
// version 1.1.1 : Support smoothing groups(#162)
// version 1.1.0 : Support parsing vertex color(#144)
// version 1.0.8 : Fix parsing `g` tag just after `usemtl`(#138)
// version 1.0.7 : Support multiple tex options(#126)
// version 1.0.6 : Add TINYOBJLOADER_USE_DOUBLE option(#124)
// version 1.0.5 : Ignore `Tr` when `d` exists in MTL(#43)
// version 1.0.4 : Support multiple filenames for 'mtllib'(#112)
// version 1.0.3 : Support parsing texture options(#85)
// version 1.0.2 : Improve parsing speed by about a factor of 2 for large
// files(#105)
// version 1.0.1 : Fixes a shape is lost if obj ends with a 'usemtl'(#104)
// version 1.0.0 : Change data structure. Change license from BSD to MIT.
//

//
// Use this in *one* .cc
//   #define TINYOBJLOADER_IMPLEMENTATION
//   #include "tiny_obj_loader.h"
//

#ifndef TINY_OBJ_LOADER_H_
#define TINY_OBJ_LOADER_H_

#include <algorithm>
#include <map>
#include <string>
#include <vector>

#include <cstdint>
#include <cstring>
#include <memory>
#include <new>
#include <type_traits>

namespace tinyobj {

// C++11 is now the minimum required standard.
#if __cplusplus < 201103L && (!defined(_MSVC_LANG) || _MSVC_LANG < 201103L)
#error "tinyobjloader requires C++11 or later. Compile with -std=c++11 or higher."
#endif

#ifdef __clang__
#pragma clang diagnostic push
#if __has_warning("-Wzero-as-null-pointer-constant")
#pragma clang diagnostic ignored "-Wzero-as-null-pointer-constant"
#endif

#pragma clang diagnostic ignored "-Wpadded"

#endif

// https://en.wikipedia.org/wiki/Wavefront_.obj_file says ...
//
//  -blendu on | off                       # set horizontal texture blending
//  (default on)
//  -blendv on | off                       # set vertical texture blending
//  (default on)
//  -boost real_value                      # boost mip-map sharpness
//  -mm base_value gain_value              # modify texture map values (default
//  0 1)
//                                         #     base_value = brightness,
//                                         gain_value = contrast
//  -o u [v [w]]                           # Origin offset             (default
//  0 0 0)
//  -s u [v [w]]                           # Scale                     (default
//  1 1 1)
//  -t u [v [w]]                           # Turbulence                (default
//  0 0 0)
//  -texres resolution                     # texture resolution to create
//  -clamp on | off                        # only render texels in the clamped
//  0-1 range (default off)
//                                         #   When unclamped, textures are
//                                         repeated across a surface,
//                                         #   when clamped, only texels which
//                                         fall within the 0-1
//                                         #   range are rendered.
//  -bm mult_value                         # bump multiplier (for bump maps
//  only)
//
//  -imfchan r | g | b | m | l | z         # specifies which channel of the file
//  is used to
//                                         # create a scalar or bump texture.
//                                         r:red, g:green,
//                                         # b:blue, m:matte, l:luminance,
//                                         z:z-depth..
//                                         # (the default for bump is 'l' and
//                                         for decal is 'm')
//  bump -imfchan r bumpmap.tga            # says to use the red channel of
//  bumpmap.tga as the bumpmap
//
// For reflection maps...
//
//   -type sphere                           # specifies a sphere for a "refl"
//   reflection map
//   -type cube_top    | cube_bottom |      # when using a cube map, the texture
//   file for each
//         cube_front  | cube_back   |      # side of the cube is specified
//         separately
//         cube_left   | cube_right
//
// TinyObjLoader extension.
//
//   -colorspace SPACE                      # Color space of the texture. e.g.
//   'sRGB` or 'linear'
//

#ifdef TINYOBJLOADER_USE_DOUBLE
//#pragma message "using double"
typedef double real_t;
#else
//#pragma message "using float"
typedef float real_t;
#endif

typedef enum {
  TEXTURE_TYPE_NONE,  // default
  TEXTURE_TYPE_SPHERE,
  TEXTURE_TYPE_CUBE_TOP,
  TEXTURE_TYPE_CUBE_BOTTOM,
  TEXTURE_TYPE_CUBE_FRONT,
  TEXTURE_TYPE_CUBE_BACK,
  TEXTURE_TYPE_CUBE_LEFT,
  TEXTURE_TYPE_CUBE_RIGHT
} texture_type_t;

struct texture_option_t {
  texture_type_t type;      // -type (default TEXTURE_TYPE_NONE)
  real_t sharpness;         // -boost (default 1.0?)
  real_t brightness;        // base_value in -mm option (default 0)
  real_t contrast;          // gain_value in -mm option (default 1)
  real_t origin_offset[3];  // -o u [v [w]] (default 0 0 0)
  real_t scale[3];          // -s u [v [w]] (default 1 1 1)
  real_t turbulence[3];     // -t u [v [w]] (default 0 0 0)
  int texture_resolution;   // -texres resolution (No default value in the spec.
                            // We'll use -1)
  bool clamp;               // -clamp (default false)
  char imfchan;  // -imfchan (the default for bump is 'l' and for decal is 'm')
  bool blendu;   // -blendu (default on)
  bool blendv;   // -blendv (default on)
  real_t bump_multiplier;  // -bm (for bump maps only, default 1.0)

  // extension
  std::string colorspace;  // Explicitly specify color space of stored texel
                           // value. Usually `sRGB` or `linear` (default empty).
};

struct material_t {
  std::string name;

  real_t ambient[3];
  real_t diffuse[3];
  real_t specular[3];
  real_t transmittance[3];
  real_t emission[3];
  real_t shininess;
  real_t ior;       // index of refraction
  real_t dissolve;  // 1 == opaque; 0 == fully transparent
  // illumination model (see http://www.fileformat.info/format/material/)
  int illum;

  int dummy;  // Suppress padding warning.

  std::string ambient_texname;   // map_Ka. For ambient or ambient occlusion.
  std::string diffuse_texname;   // map_Kd
  std::string specular_texname;  // map_Ks
  std::string specular_highlight_texname;  // map_Ns
  std::string bump_texname;                // map_bump, map_Bump, bump
  std::string displacement_texname;        // disp
  std::string alpha_texname;               // map_d
  std::string reflection_texname;          // refl

  texture_option_t ambient_texopt;
  texture_option_t diffuse_texopt;
  texture_option_t specular_texopt;
  texture_option_t specular_highlight_texopt;
  texture_option_t bump_texopt;
  texture_option_t displacement_texopt;
  texture_option_t alpha_texopt;
  texture_option_t reflection_texopt;

  // PBR extension
  // http://exocortex.com/blog/extending_wavefront_mtl_to_support_pbr
  real_t roughness;            // [0, 1] default 0
  real_t metallic;             // [0, 1] default 0
  real_t sheen;                // [0, 1] default 0
  real_t clearcoat_thickness;  // [0, 1] default 0
  real_t clearcoat_roughness;  // [0, 1] default 0
  real_t anisotropy;           // aniso. [0, 1] default 0
  real_t anisotropy_rotation;  // anisor. [0, 1] default 0
  real_t pad0;
  std::string roughness_texname;  // map_Pr
  std::string metallic_texname;   // map_Pm
  std::string sheen_texname;      // map_Ps
  std::string emissive_texname;   // map_Ke
  std::string normal_texname;     // norm. For normal mapping.

  texture_option_t roughness_texopt;
  texture_option_t metallic_texopt;
  texture_option_t sheen_texopt;
  texture_option_t emissive_texopt;
  texture_option_t normal_texopt;

  int pad2;

  std::map<std::string, std::string> unknown_parameter;

#ifdef TINY_OBJ_LOADER_PYTHON_BINDING
  // For pybind11
  std::array<double, 3> GetDiffuse() {
    std::array<double, 3> values;
    values[0] = double(diffuse[0]);
    values[1] = double(diffuse[1]);
    values[2] = double(diffuse[2]);

    return values;
  }

  std::array<double, 3> GetSpecular() {
    std::array<double, 3> values;
    values[0] = double(specular[0]);
    values[1] = double(specular[1]);
    values[2] = double(specular[2]);

    return values;
  }

  std::array<double, 3> GetTransmittance() {
    std::array<double, 3> values;
    values[0] = double(transmittance[0]);
    values[1] = double(transmittance[1]);
    values[2] = double(transmittance[2]);

    return values;
  }

  std::array<double, 3> GetEmission() {
    std::array<double, 3> values;
    values[0] = double(emission[0]);
    values[1] = double(emission[1]);
    values[2] = double(emission[2]);

    return values;
  }

  std::array<double, 3> GetAmbient() {
    std::array<double, 3> values;
    values[0] = double(ambient[0]);
    values[1] = double(ambient[1]);
    values[2] = double(ambient[2]);

    return values;
  }

  void SetDiffuse(std::array<double, 3> &a) {
    diffuse[0] = real_t(a[0]);
    diffuse[1] = real_t(a[1]);
    diffuse[2] = real_t(a[2]);
  }

  void SetAmbient(std::array<double, 3> &a) {
    ambient[0] = real_t(a[0]);
    ambient[1] = real_t(a[1]);
    ambient[2] = real_t(a[2]);
  }

  void SetSpecular(std::array<double, 3> &a) {
    specular[0] = real_t(a[0]);
    specular[1] = real_t(a[1]);
    specular[2] = real_t(a[2]);
  }

  void SetTransmittance(std::array<double, 3> &a) {
    transmittance[0] = real_t(a[0]);
    transmittance[1] = real_t(a[1]);
    transmittance[2] = real_t(a[2]);
  }

  std::string GetCustomParameter(const std::string &key) {
    std::map<std::string, std::string>::const_iterator it =
        unknown_parameter.find(key);

    if (it != unknown_parameter.end()) {
      return it->second;
    }
    return std::string();
  }

#endif
};

template <typename Alloc = std::allocator<char>>
struct basic_tag_t {
  using allocator_type = Alloc;
  using char_alloc =
      typename std::allocator_traits<Alloc>::template rebind_alloc<char>;
  using string_type =
      std::basic_string<char, std::char_traits<char>, char_alloc>;
  using int_alloc =
      typename std::allocator_traits<Alloc>::template rebind_alloc<int>;
  using real_alloc =
      typename std::allocator_traits<Alloc>::template rebind_alloc<real_t>;
  using string_alloc =
      typename std::allocator_traits<Alloc>::template rebind_alloc<string_type>;

  string_type name;
  std::vector<int, int_alloc> intValues;
  std::vector<real_t, real_alloc> floatValues;
  std::vector<string_type, string_alloc> stringValues;

  basic_tag_t()
      : name(),
        intValues(),
        floatValues(),
        stringValues() {}

  explicit basic_tag_t(const allocator_type &alloc)
      : name(char_alloc(alloc)),
        intValues(int_alloc(alloc)),
        floatValues(real_alloc(alloc)),
        stringValues(string_alloc(alloc)) {}

  template <typename OtherAlloc>
  basic_tag_t(const basic_tag_t<OtherAlloc> &rhs) : name(rhs.name) {
    intValues.assign(rhs.intValues.begin(), rhs.intValues.end());
    floatValues.assign(rhs.floatValues.begin(), rhs.floatValues.end());
    stringValues.assign(rhs.stringValues.begin(), rhs.stringValues.end());
  }

  template <typename OtherAlloc>
  basic_tag_t(const basic_tag_t<OtherAlloc> &rhs, const allocator_type &alloc)
      : name(rhs.name.begin(), rhs.name.end(), char_alloc(alloc)),
        intValues(int_alloc(alloc)),
        floatValues(real_alloc(alloc)),
        stringValues(string_alloc(alloc)) {
    intValues.assign(rhs.intValues.begin(), rhs.intValues.end());
    floatValues.assign(rhs.floatValues.begin(), rhs.floatValues.end());
    for (size_t i = 0; i < rhs.stringValues.size(); i++) {
      stringValues.emplace_back(rhs.stringValues[i].begin(),
                                rhs.stringValues[i].end(), char_alloc(alloc));
    }
  }

  template <typename OtherAlloc>
  basic_tag_t &operator=(const basic_tag_t<OtherAlloc> &rhs) {
    name = rhs.name;
    intValues.assign(rhs.intValues.begin(), rhs.intValues.end());
    floatValues.assign(rhs.floatValues.begin(), rhs.floatValues.end());
    stringValues.assign(rhs.stringValues.begin(), rhs.stringValues.end());
    return *this;
  }
};

using tag_t = basic_tag_t<>;

struct joint_and_weight_t {
  int joint_id;
  real_t weight;
};

template <typename Alloc = std::allocator<char>>
struct basic_skin_weight_t {
  using allocator_type = Alloc;
  using joint_weight_alloc =
      typename std::allocator_traits<Alloc>::template rebind_alloc<
          joint_and_weight_t>;

  int vertex_id;  // Corresponding vertex index in `attrib_t::vertices`.
                  // Compared to `index_t`, this index must be positive and
                  // start with 0(does not allow relative indexing)
  std::vector<joint_and_weight_t, joint_weight_alloc> weightValues;

  basic_skin_weight_t() : vertex_id(0) {}

  explicit basic_skin_weight_t(const allocator_type &alloc)
      : vertex_id(0), weightValues(joint_weight_alloc(alloc)) {}

  template <typename OtherAlloc>
  basic_skin_weight_t(const basic_skin_weight_t<OtherAlloc> &rhs)
      : vertex_id(rhs.vertex_id) {
    weightValues.assign(rhs.weightValues.begin(), rhs.weightValues.end());
  }

  template <typename OtherAlloc>
  basic_skin_weight_t(const basic_skin_weight_t<OtherAlloc> &rhs,
                      const allocator_type &alloc)
      : vertex_id(rhs.vertex_id), weightValues(joint_weight_alloc(alloc)) {
    weightValues.assign(rhs.weightValues.begin(), rhs.weightValues.end());
  }

  template <typename OtherAlloc>
  basic_skin_weight_t &operator=(const basic_skin_weight_t<OtherAlloc> &rhs) {
    vertex_id = rhs.vertex_id;
    weightValues.assign(rhs.weightValues.begin(), rhs.weightValues.end());
    return *this;
  }
};

using skin_weight_t = basic_skin_weight_t<>;

// Index struct to support different indices for vtx/normal/texcoord.
// -1 means not used.
struct index_t {
  int vertex_index;
  int normal_index;
  int texcoord_index;
};

struct mesh_t {
  std::vector<index_t> indices;
  std::vector<unsigned int>
      num_face_vertices;          // The number of vertices per
                                  // face. 3 = triangle, 4 = quad, ...
  std::vector<int> material_ids;  // per-face material ID
  std::vector<unsigned int> smoothing_group_ids;  // per-face smoothing group
                                                  // ID(0 = off. positive value
                                                  // = group id)
  std::vector<tag_t> tags;                        // SubD tag
};

// struct path_t {
//  std::vector<int> indices;  // pairs of indices for lines
//};

struct lines_t {
  // Linear flattened indices.
  std::vector<index_t> indices;        // indices for vertices(poly lines)
  std::vector<int> num_line_vertices;  // The number of vertices per line.
};

struct points_t {
  std::vector<index_t> indices;  // indices for points
};

struct shape_t {
  std::string name;
  mesh_t mesh;
  lines_t lines;
  points_t points;
};

// Vertex attributes
struct attrib_t {
  std::vector<real_t> vertices;  // 'v'(xyz)

  // For backward compatibility, we store vertex weight in separate array.
  std::vector<real_t> vertex_weights;  // 'v'(w)
  std::vector<real_t> normals;         // 'vn'
  std::vector<real_t> texcoords;       // 'vt'(uv)

  // For backward compatibility, we store texture coordinate 'w' in separate
  // array.
  std::vector<real_t> texcoord_ws;  // 'vt'(w)
  std::vector<real_t> colors;       // extension: vertex colors

  //
  // TinyObj extension.
  //

  // NOTE(syoyo): array index is based on the appearance order.
  // To get a corresponding skin weight for a specific vertex id `vid`,
  // Need to reconstruct a look up table: `skin_weight_t::vertex_id` == `vid`
  // (e.g. using std::map, std::unordered_map)
  std::vector<skin_weight_t> skin_weights;

  attrib_t() {}

  //
  // For pybind11
  //
  const std::vector<real_t> &GetVertices() const { return vertices; }

  const std::vector<real_t> &GetVertexWeights() const { return vertex_weights; }
};

struct callback_t {
  // W is optional and set to 1 if there is no `w` item in `v` line
  void (*vertex_cb)(void *user_data, real_t x, real_t y, real_t z, real_t w);
  void (*vertex_color_cb)(void *user_data, real_t x, real_t y, real_t z,
                          real_t r, real_t g, real_t b, bool has_color);
  void (*normal_cb)(void *user_data, real_t x, real_t y, real_t z);

  // y and z are optional and set to 0 if there is no `y` and/or `z` item(s) in
  // `vt` line.
  void (*texcoord_cb)(void *user_data, real_t x, real_t y, real_t z);

  // called per 'f' line. num_indices is the number of face indices(e.g. 3 for
  // triangle, 4 for quad)
  // 0 will be passed for undefined index in index_t members.
  void (*index_cb)(void *user_data, index_t *indices, int num_indices);
  // `name` material name, `material_id` = the array index of material_t[]. -1
  // if
  // a material not found in .mtl
  void (*usemtl_cb)(void *user_data, const char *name, int material_id);
  // `materials` = parsed material data.
  void (*mtllib_cb)(void *user_data, const material_t *materials,
                    int num_materials);
  // There may be multiple group names
  void (*group_cb)(void *user_data, const char **names, int num_names);
  void (*object_cb)(void *user_data, const char *name);

  callback_t()
      : vertex_cb(NULL),
        vertex_color_cb(NULL),
        normal_cb(NULL),
        texcoord_cb(NULL),
        index_cb(NULL),
        usemtl_cb(NULL),
        mtllib_cb(NULL),
        group_cb(NULL),
        object_cb(NULL) {}
};

class MaterialReader {
 public:
  MaterialReader() {}
  virtual ~MaterialReader();

  virtual bool operator()(const std::string &matId,
                          std::vector<material_t> *materials,
                          std::map<std::string, int> *matMap, std::string *warn,
                          std::string *err) = 0;
};

///
/// Read .mtl from a file.
///
class MaterialFileReader : public MaterialReader {
 public:
  // Path could contain separator(';' in Windows, ':' in Posix)
  explicit MaterialFileReader(const std::string &mtl_basedir)
      : m_mtlBaseDir(mtl_basedir) {}
  virtual ~MaterialFileReader() override {}
  virtual bool operator()(const std::string &matId,
                          std::vector<material_t> *materials,
                          std::map<std::string, int> *matMap, std::string *warn,
                          std::string *err) override;

 private:
  std::string m_mtlBaseDir;
};

///
/// Read .mtl from a stream.
///
class MaterialStreamReader : public MaterialReader {
 public:
  explicit MaterialStreamReader(std::istream &inStream)
      : m_inStream(inStream) {}
  virtual ~MaterialStreamReader() override {}
  virtual bool operator()(const std::string &matId,
                          std::vector<material_t> *materials,
                          std::map<std::string, int> *matMap, std::string *warn,
                          std::string *err) override;

 private:
  std::istream &m_inStream;
};

// v2 API
struct ObjReaderConfig {
  bool triangulate;  // triangulate polygon?

  // Currently not used.
  // "simple" or empty: Create triangle fan
  // "earcut": Use the algorithm based on Ear clipping
  std::string triangulation_method;

  /// Parse vertex color.
  /// If vertex color is not present, its filled with default value.
  /// false = no vertex color
  /// This will increase memory of parsed .obj
  bool vertex_color;

  ///
  /// Search path to .mtl file.
  /// Default = "" = search from the same directory of .obj file.
  /// Valid only when loading .obj from a file.
  ///
  std::string mtl_search_path;

  ObjReaderConfig()
      : triangulate(true), triangulation_method("simple"), vertex_color(true) {}
};

///
/// Wavefront .obj reader class(v2 API)
///
class ObjReader {
 public:
  ObjReader() : valid_(false) {}

  ///
  /// Load .obj and .mtl from a file.
  ///
  /// @param[in] filename wavefront .obj filename
  /// @param[in] config Reader configuration
  ///
  bool ParseFromFile(const std::string &filename,
                     const ObjReaderConfig &config = ObjReaderConfig());

  ///
  /// Parse .obj from a text string.
  /// Need to supply .mtl text string by `mtl_text`.
  /// This function ignores `mtllib` line in .obj text.
  ///
  /// @param[in] obj_text wavefront .obj filename
  /// @param[in] mtl_text wavefront .mtl filename
  /// @param[in] config Reader configuration
  ///
  bool ParseFromString(const std::string &obj_text, const std::string &mtl_text,
                       const ObjReaderConfig &config = ObjReaderConfig());

  ///
  /// .obj was loaded or parsed correctly.
  ///
  bool Valid() const { return valid_; }

  const attrib_t &GetAttrib() const { return attrib_; }

  const std::vector<shape_t> &GetShapes() const { return shapes_; }

  const std::vector<material_t> &GetMaterials() const { return materials_; }

  ///
  /// Warning message(may be filled after `Load` or `Parse`)
  ///
  const std::string &Warning() const { return warning_; }

  ///
  /// Error message(filled when `Load` or `Parse` failed)
  ///
  const std::string &Error() const { return error_; }

 private:
  bool valid_;

  attrib_t attrib_;
  std::vector<shape_t> shapes_;
  std::vector<material_t> materials_;

  std::string warning_;
  std::string error_;
};

/// ==>>========= Legacy v1 API =============================================

/// Loads .obj from a file.
/// 'attrib', 'shapes' and 'materials' will be filled with parsed shape data
/// 'shapes' will be filled with parsed shape data
/// Returns true when loading .obj become success.
/// Returns warning message into `warn`, and error message into `err`
/// 'mtl_basedir' is optional, and used for base directory for .mtl file.
/// In default(`NULL'), .mtl file is searched from an application's working
/// directory.
/// 'triangulate' is optional, and used whether triangulate polygon face in .obj
/// or not.
/// Option 'default_vcols_fallback' specifies whether vertex colors should
/// always be defined, even if no colors are given (fallback to white).
bool LoadObj(attrib_t *attrib, std::vector<shape_t> *shapes,
             std::vector<material_t> *materials, std::string *warn,
             std::string *err, const char *filename,
             const char *mtl_basedir = NULL, bool triangulate = true,
             bool default_vcols_fallback = true);

/// Loads .obj from a file with custom user callback.
/// .mtl is loaded as usual and parsed material_t data will be passed to
/// `callback.mtllib_cb`.
/// Returns true when loading .obj/.mtl become success.
/// Returns warning message into `warn`, and error message into `err`
/// See `examples/callback_api/` for how to use this function.
bool LoadObjWithCallback(std::istream &inStream, const callback_t &callback,
                         void *user_data = NULL,
                         MaterialReader *readMatFn = NULL,
                         std::string *warn = NULL, std::string *err = NULL);

/// Loads object from a std::istream, uses `readMatFn` to retrieve
/// std::istream for materials.
/// Returns true when loading .obj become success.
/// Returns warning and error message into `err`
bool LoadObj(attrib_t *attrib, std::vector<shape_t> *shapes,
             std::vector<material_t> *materials, std::string *warn,
             std::string *err, std::istream *inStream,
             MaterialReader *readMatFn = NULL, bool triangulate = true,
             bool default_vcols_fallback = true);

/// Loads materials into std::map
void LoadMtl(std::map<std::string, int> *material_map,
             std::vector<material_t> *materials, std::istream *inStream,
             std::string *warning, std::string *err);

///
/// Parse texture name and texture option for custom texture parameter through
/// material::unknown_parameter
///
/// @param[out] texname Parsed texture name
/// @param[out] texopt Parsed texopt
/// @param[in] linebuf Input string
///
bool ParseTextureNameAndOption(std::string *texname, texture_option_t *texopt,
                               const char *linebuf);

/// =<<========== Legacy v1 API =============================================

/// ==>>========= Optimized API (C++11 required) ============================
///
/// Optional compile options (all DISABLED by default — define the macro to
/// opt in):
///   TINYOBJLOADER_USE_MULTITHREADING - multi-threaded parsing
///   TINYOBJLOADER_USE_SIMD           - SIMD-accelerated line scanning
///                                      (SSE2/AVX2 on x86, NEON on ARM)
///   TINYOBJLOADER_ENABLE_EXCEPTION   - throw std::bad_alloc on allocation
///                                      failure instead of returning nullptr
///
/// With none of these defined, the parser runs single-threaded, scalar, and
/// exception-free.  These features require C++11 or later.
///

///
/// Arena-based memory allocator for reduced allocation overhead when
/// loading huge meshes.  Memory is freed in bulk when the arena is
/// destroyed or reset().  Individual deallocate() calls are no-ops.
///
class ArenaAllocator {
 public:
  explicit ArenaAllocator(size_t block_size = 1024 * 1024)
      : head_(nullptr), default_block_size_(block_size) {}

  ~ArenaAllocator() { destroy(); }

  ArenaAllocator(const ArenaAllocator &) = delete;
  ArenaAllocator &operator=(const ArenaAllocator &) = delete;

  ArenaAllocator(ArenaAllocator &&o) noexcept
      : head_(o.head_), default_block_size_(o.default_block_size_) {
    o.head_ = nullptr;
  }
  ArenaAllocator &operator=(ArenaAllocator &&o) noexcept {
    if (this != &o) {
      destroy();
      head_ = o.head_;
      default_block_size_ = o.default_block_size_;
      o.head_ = nullptr;
    }
    return *this;
  }

  void *allocate(size_t bytes,
                 size_t alignment = sizeof(void *));

  /// Free all memory at once.
  void reset();

 private:
  struct Block {
    unsigned char *data;
    size_t capacity;
    size_t used;
    Block *next;
  };

  Block *head_;
  size_t default_block_size_;

  Block *new_block(size_t min_bytes);
  void destroy();
};

///
/// STL-compatible allocator adapter backed by an ArenaAllocator.
/// deallocate() is a no-op — memory is released when the arena is reset.
///
template <typename T>
class arena_adapter {
 public:
  using value_type = T;
  using pointer = T *;
  using const_pointer = const T *;
  using size_type = std::size_t;
  using difference_type = std::ptrdiff_t;
  using propagate_on_container_copy_assignment = std::true_type;
  using propagate_on_container_move_assignment = std::true_type;
  using propagate_on_container_swap = std::true_type;

  explicit arena_adapter(ArenaAllocator *arena = nullptr) noexcept
      : arena_(arena) {}

  template <typename U>
  arena_adapter(const arena_adapter<U> &other) noexcept
      : arena_(other.arena()) {}

  T *allocate(size_t n) {
    if (n > SIZE_MAX / sizeof(T)) {
#ifdef TINYOBJLOADER_ENABLE_EXCEPTION
      throw std::bad_alloc();
#else
      return nullptr;
#endif
    }
    if (arena_) {
      return static_cast<T *>(arena_->allocate(n * sizeof(T), alignof(T)));
    }
#ifdef TINYOBJLOADER_ENABLE_EXCEPTION
    return static_cast<T *>(::operator new(n * sizeof(T)));
#else
    return static_cast<T *>(::operator new(n * sizeof(T), std::nothrow));
#endif
  }

  void deallocate(T *p, size_t) noexcept {
    if (!arena_) {
      ::operator delete(p);
      return;
    }
    // Arena deallocation is a no-op; memory freed in bulk via reset().
  }

  ArenaAllocator *arena() const noexcept { return arena_; }

  template <typename U>
  bool operator==(const arena_adapter<U> &o) const noexcept {
    return arena_ == o.arena();
  }
  template <typename U>
  bool operator!=(const arena_adapter<U> &o) const noexcept {
    return arena_ != o.arena();
  }

  template <typename U>
  struct rebind {
    using other = arena_adapter<U>;
  };

 private:
  ArenaAllocator *arena_;
};

///
/// Template mesh type supporting custom allocators.
/// Note: Alloc must be default-constructible (stateless). For stateful
/// allocators like arena_adapter, construct vectors individually with
/// an allocator instance.
///
template <typename Alloc = std::allocator<char>>
struct basic_mesh_t {
  using index_alloc =
      typename std::allocator_traits<Alloc>::template rebind_alloc<index_t>;
  using uint_alloc =
      typename std::allocator_traits<Alloc>::template rebind_alloc<unsigned int>;
  using int_alloc =
      typename std::allocator_traits<Alloc>::template rebind_alloc<int>;
  using tag_alloc =
      typename std::allocator_traits<Alloc>::template rebind_alloc<
          basic_tag_t<Alloc> >;

  std::vector<index_t, index_alloc> indices;
  std::vector<unsigned int, uint_alloc> num_face_vertices;
  std::vector<int, int_alloc> material_ids;
  std::vector<unsigned int, uint_alloc> smoothing_group_ids;
  std::vector<basic_tag_t<Alloc>, tag_alloc> tags;
};

template <typename Alloc = std::allocator<char>>
struct basic_lines_t {
  using index_alloc =
      typename std::allocator_traits<Alloc>::template rebind_alloc<index_t>;
  using int_alloc =
      typename std::allocator_traits<Alloc>::template rebind_alloc<int>;

  std::vector<index_t, index_alloc> indices;
  std::vector<int, int_alloc> num_line_vertices;
};

template <typename Alloc = std::allocator<char>>
struct basic_points_t {
  using index_alloc =
      typename std::allocator_traits<Alloc>::template rebind_alloc<index_t>;

  std::vector<index_t, index_alloc> indices;
};

///
/// Template shape type supporting custom allocators.
/// `name` always uses the default allocator; only mesh buffers are
/// allocator-aware.
///
template <typename Alloc = std::allocator<char>>
struct basic_shape_t {
  std::string name;
  basic_mesh_t<Alloc> mesh;
  basic_lines_t<Alloc> lines;
  basic_points_t<Alloc> points;
};

///
/// Template attrib type supporting custom allocators.
/// Flat arrays: vertices(xyz), normals(xyz), texcoords(uv).
/// Note: Alloc must be default-constructible (stateless). For stateful
/// allocators like arena_adapter, construct vectors individually with
/// an allocator instance.
///
template <typename Alloc = std::allocator<char>>
struct basic_attrib_t {
  using real_alloc =
      typename std::allocator_traits<Alloc>::template rebind_alloc<real_t>;
  using int_alloc =
      typename std::allocator_traits<Alloc>::template rebind_alloc<int>;
  using index_alloc =
      typename std::allocator_traits<Alloc>::template rebind_alloc<index_t>;
  using skin_weight_alloc =
      typename std::allocator_traits<Alloc>::template rebind_alloc<
          basic_skin_weight_t<Alloc> >;

  std::vector<real_t, real_alloc> vertices;   // xyz
  std::vector<real_t, real_alloc> vertex_weights;  // optional w for `v`
  std::vector<real_t, real_alloc> normals;    // xyz
  std::vector<real_t, real_alloc> texcoords;  // uv
  std::vector<real_t, real_alloc> texcoord_ws;  // optional w for `vt`
  std::vector<real_t, real_alloc> colors;     // rgb (optional)
  std::vector<basic_skin_weight_t<Alloc>, skin_weight_alloc> skin_weights;
  std::vector<index_t, index_alloc> indices;  // face indices
  std::vector<int, int_alloc> face_num_verts; // verts per face
  std::vector<int, int_alloc> material_ids;   // per-face material
};

///
/// Configuration for the optimized loader.
///
struct OptLoadConfig {
  /// Number of threads.  -1 = hardware_concurrency, 0 or 1 = single-threaded.
  /// Effective only when TINYOBJLOADER_USE_MULTITHREADING is defined.
  int num_threads;

  bool triangulate;  ///< Triangulate polygons (fan triangulation).
  bool verbose;      ///< Reserved for future use (currently has no effect).

  /// Enable trie-based float string → value cache.
  /// Speeds up files with many repeated coordinate values (e.g. "0.0", "1.0").
  bool float_cache;

  /// Maximum trie nodes per thread (controls cache memory).
  /// Each node is ~36 bytes, so 1024 nodes ≈ 36 KB (fits in L1 cache).
  int float_cache_max_nodes;

  /// Use fp32-length keys for the float cache (max 8 chars).
  /// When true, only tokens up to `float::digits10 + 2` characters are cached.
  /// When false, uses `real_t::digits10 + 2` (longer keys, fewer hits).
  /// Has no effect when float_cache is false.
  bool fp32_cache;

  OptLoadConfig()
      : num_threads(-1), triangulate(true), verbose(false),
        float_cache(false), float_cache_max_nodes(1024), fp32_cache(true) {}
};

/// Optimized loader — parse from a raw memory buffer.
/// Supports multi-threading (TINYOBJLOADER_USE_MULTITHREADING) and
/// SIMD line scanning (TINYOBJLOADER_USE_SIMD).
bool LoadObjOpt(basic_attrib_t<> *attrib,
                std::vector<basic_shape_t<>> *shapes,
                std::vector<material_t> *materials,
                std::string *warn, std::string *err,
                const char *buf, size_t buf_len,
                const OptLoadConfig &config = OptLoadConfig());

/// Optimized loader — load from a file.
bool LoadObjOpt(basic_attrib_t<> *attrib,
                std::vector<basic_shape_t<>> *shapes,
                std::vector<material_t> *materials,
                std::string *warn, std::string *err,
                const char *filename,
                const char *mtl_basedir = nullptr,
                const OptLoadConfig &config = OptLoadConfig());

// ---- TypedArray-based zero-copy API for maximum efficiency ----

///
/// Non-owning, contiguous array view allocated from an ArenaAllocator.
/// No capacity tracking, no realloc.  Lifetime is tied to the arena.
///
template <typename T>
class TypedArray {
  static_assert(std::is_trivially_copyable<T>::value,
                "TypedArray<T> requires T to be trivially copyable "
                "(uses memset for initialization).");

 public:
  TypedArray() : data_(nullptr), size_(0) {}

  T *data() { return data_; }
  const T *data() const { return data_; }
  size_t size() const { return size_; }
  bool empty() const { return size_ == 0; }

  T &operator[](size_t i) { return data_[i]; }
  const T &operator[](size_t i) const { return data_[i]; }

  T *begin() { return data_; }
  T *end() { return data_ + size_; }
  const T *begin() const { return data_; }
  const T *end() const { return data_ + size_; }

  /// Allocate `count` elements from `arena`.  Previous contents are abandoned.
  /// Elements are zero-initialized via memset (T must be trivially copyable).
  /// Returns false if allocation fails (only possible when exceptions are
  /// disabled via TINYOBJLOADER_ENABLE_EXCEPTION not being defined).
  bool allocate(ArenaAllocator &arena, size_t count) {
    if (count == 0) {
      data_ = nullptr;
      size_ = 0;
      return true;
    }
    // Guard against size_t overflow in count * sizeof(T).
    if (count > SIZE_MAX / sizeof(T)) {
#ifdef TINYOBJLOADER_ENABLE_EXCEPTION
      throw std::bad_alloc();
#else
      data_ = nullptr;
      size_ = 0;
      return false;
#endif
    }
    void *p = arena.allocate(count * sizeof(T), alignof(T));
    if (!p) {
      data_ = nullptr;
      size_ = 0;
      return false;
    }
    data_ = static_cast<T *>(p);
    size_ = count;
    std::memset(data_, 0, count * sizeof(T));
    return true;
  }

  /// Wrap an existing arena-allocated pointer.  Caller must ensure ptr
  /// points into a live arena and count is within bounds.
  void set(T *ptr, size_t count) {
    data_ = ptr;
    size_ = count;
  }

  /// Shrink the logical size (no memory freed — arena owns it).
  void truncate(size_t new_size) {
    if (new_size < size_) size_ = new_size;
  }

  void clear() {
    data_ = nullptr;
    size_ = 0;
  }

 private:
  T *data_;
  size_t size_;
};

///
/// A shape expressed as a range into the flat attrib arrays.
/// No owned memory — just offsets + counts.
///
struct OptShapeRange {
  std::string name;
  size_t face_offset;       ///< Start index into OptAttrib::face_num_verts / material_ids
  size_t face_count;        ///< Number of faces in this shape
  size_t index_offset;      ///< Start index into OptAttrib::indices
  size_t index_count;       ///< Number of index_t entries for this shape

  OptShapeRange()
      : face_offset(0), face_count(0), index_offset(0), index_count(0) {}
};

///
/// Flat attribute storage using TypedArray (arena-backed).
/// All arrays are allocated from the OptResult's arena.
///
/// Unlike basic_attrib_t (used by LoadObjOpt), optional arrays like
/// vertex_weights, texcoord_ws, colors, and smoothing_group_ids are only
/// allocated when the input actually contains the corresponding data.
/// Check .empty() before accessing.
///
struct OptAttrib {
  TypedArray<real_t> vertices;        ///< xyz, length = num_vertices * 3
  TypedArray<real_t> vertex_weights;  ///< w per vertex (empty if no `v` line has w)
  TypedArray<real_t> normals;         ///< xyz, length = num_normals * 3
  TypedArray<real_t> texcoords;       ///< uv, length = num_texcoords * 2
  TypedArray<real_t> texcoord_ws;     ///< w per texcoord (empty if no `vt` has w)
  TypedArray<real_t> colors;          ///< rgb per vertex (empty if not all verts have color)

  TypedArray<index_t> indices;        ///< flattened face indices
  TypedArray<int> face_num_verts;     ///< vertices per face
  TypedArray<int> material_ids;       ///< per-face material id
  TypedArray<unsigned int> smoothing_group_ids; ///< per-face smoothing group
};

///
/// Complete result from the TypedArray-based optimized loader.
/// Owns the ArenaAllocator — all TypedArray pointers are valid as long
/// as this object is alive.  Move-only.
///
struct OptResult {
  ArenaAllocator arena;
  OptAttrib attrib;
  std::vector<OptShapeRange> shapes;  ///< views into attrib arrays
  std::vector<material_t> materials;
  bool valid;

  OptResult() : arena(4 * 1024 * 1024), valid(false) {}

  OptResult(OptResult &&o) noexcept
      : arena(std::move(o.arena)),
        attrib(o.attrib),
        shapes(std::move(o.shapes)),
        materials(std::move(o.materials)),
        valid(o.valid) {
    o.attrib = OptAttrib();  // clear dangling pointers
    o.valid = false;
  }

  OptResult &operator=(OptResult &&o) noexcept {
    if (this != &o) {
      arena = std::move(o.arena);
      attrib = o.attrib;
      shapes = std::move(o.shapes);
      materials = std::move(o.materials);
      valid = o.valid;
      o.attrib = OptAttrib();
      o.valid = false;
    }
    return *this;
  }

  OptResult(const OptResult &) = delete;
  OptResult &operator=(const OptResult &) = delete;
};

/// TypedArray-based optimized loader — parse from a raw memory buffer.
/// Returns OptResult that owns all allocated memory via its arena.
OptResult LoadObjOptTyped(const char *buf, size_t buf_len,
                          std::string *warn, std::string *err,
                          const OptLoadConfig &config = OptLoadConfig());

/// TypedArray-based optimized loader — load from a file.
OptResult LoadObjOptTyped(const char *filename,
                          std::string *warn, std::string *err,
                          const char *mtl_basedir = nullptr,
                          const OptLoadConfig &config = OptLoadConfig());

/// =<<========== Optimized API =============================================

}  // namespace tinyobj

#endif  // TINY_OBJ_LOADER_H_

// Guard the implementation against double inclusion within a single
// translation unit (e.g. when another header that also `#include`s this file
// is pulled in after TINYOBJLOADER_IMPLEMENTATION is defined).
#if defined(TINYOBJLOADER_IMPLEMENTATION) && \
    !defined(TINYOBJLOADER_IMPLEMENTATION_DEFINED)
#define TINYOBJLOADER_IMPLEMENTATION_DEFINED
#include <cassert>
#include <cctype>
#include <climits>
#include <cmath>
#include <cstddef>
#include <cstdint>
#include <cerrno>
#include <cstdlib>
#include <cstring>
#include <fstream>
#include <limits>

#ifdef _WIN32
#ifndef WIN32_LEAN_AND_MEAN
#define WIN32_LEAN_AND_MEAN
#endif
#ifndef NOMINMAX
#define NOMINMAX
#endif
#include <windows.h>
#endif

#ifdef TINYOBJLOADER_USE_MMAP
#if !defined(_WIN32)
// POSIX headers for mmap
#include <fcntl.h>
#include <sys/mman.h>
#include <sys/stat.h>
#include <unistd.h>
#endif
#endif  // TINYOBJLOADER_USE_MMAP
#include <set>
#include <sstream>
#include <utility>

#ifdef TINYOBJLOADER_USE_MAPBOX_EARCUT

#ifdef TINYOBJLOADER_DONOT_INCLUDE_MAPBOX_EARCUT
// Assume earcut.hpp is included outside of tiny_obj_loader.h
#else

#ifdef __clang__
#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Weverything"
#endif

#include <array>

#include "mapbox/earcut.hpp"

#ifdef __clang__
#pragma clang diagnostic pop
#endif

#endif

#endif  // TINYOBJLOADER_USE_MAPBOX_EARCUT

#ifdef _WIN32
// Converts a UTF-8 encoded string to a UTF-16 wide string for use with
// Windows file APIs that support Unicode paths (including paths longer than
// MAX_PATH when combined with the extended-length path prefix).
static std::wstring UTF8ToWchar(const std::string &str) {
  if (str.empty()) return std::wstring();
  int size_needed =
      MultiByteToWideChar(CP_UTF8, 0, str.c_str(),
                          static_cast<int>(str.size()), NULL, 0);
  if (size_needed == 0) return std::wstring();
  std::wstring wstr(static_cast<size_t>(size_needed), L'\0');
  int result =
      MultiByteToWideChar(CP_UTF8, 0, str.c_str(),
                          static_cast<int>(str.size()), &wstr[0], size_needed);
  if (result == 0) return std::wstring();
  return wstr;
}

// Prepends the Windows extended-length path prefix ("\\?\") to an absolute
// path when the path length meets or exceeds MAX_PATH (260 characters).
// This allows Windows APIs to handle paths up to 32767 characters long.
// UNC paths (starting with "\\") are converted to "\\?\UNC\" form.
static std::wstring LongPathW(const std::wstring &wpath) {
  const std::wstring kLongPathPrefix = L"\\\\?\\";
  const std::wstring kUNCPrefix = L"\\\\";
  const std::wstring kLongUNCPathPrefix = L"\\\\?\\UNC\\";

  // Already has the extended-length prefix; return as-is.
  if (wpath.size() >= kLongPathPrefix.size() &&
      wpath.substr(0, kLongPathPrefix.size()) == kLongPathPrefix) {
    return wpath;
  }

  // Only add the prefix when the path is long enough to require it.
  if (wpath.size() < MAX_PATH) {
    return wpath;
  }

  // Normalize forward slashes to backslashes: the extended-length "\\?\"
  // prefix requires backslash separators only.
  std::wstring normalized = wpath;
  for (std::wstring::size_type i = 0; i < normalized.size(); ++i) {
    if (normalized[i] == L'/') normalized[i] = L'\\';
  }

  // UNC path: "\\server\share\..." -> "\\?\UNC\server\share\..."
  if (normalized.size() >= kUNCPrefix.size() &&
      normalized.substr(0, kUNCPrefix.size()) == kUNCPrefix) {
    return kLongUNCPathPrefix + normalized.substr(kUNCPrefix.size());
  }

  // Absolute path with drive letter: "C:\..." -> "\\?\C:\..."
  if (normalized.size() >= 2 && normalized[1] == L':') {
    return kLongPathPrefix + normalized;
  }

  return normalized;
}
#endif  // _WIN32

#ifdef TINYOBJLOADER_USE_MULTITHREADING
#include <atomic>
#include <thread>
#endif
#ifdef TINYOBJLOADER_USE_SIMD
#if defined(__SSE2__) || defined(_M_X64) || defined(_M_AMD64) || \
    (defined(_M_IX86_FP) && _M_IX86_FP >= 2)
#define TINYOBJLOADER_SIMD_SSE2 1
#include <emmintrin.h>
#if defined(__AVX2__)
#define TINYOBJLOADER_SIMD_AVX2 1
#include <immintrin.h>
#endif
#elif defined(__ARM_NEON) || defined(__ARM_NEON__)
#define TINYOBJLOADER_SIMD_NEON 1
#include <arm_neon.h>
#endif
#endif  // TINYOBJLOADER_USE_SIMD

// --------------------------------------------------------------------------
// Embedded fast_float v8.0.2 for high-performance, bit-exact float parsing.
// Disable by defining TINYOBJLOADER_DISABLE_FAST_FLOAT before including
// this file with TINYOBJLOADER_IMPLEMENTATION.
// --------------------------------------------------------------------------
#ifndef TINYOBJLOADER_DISABLE_FAST_FLOAT

// Standard headers needed by the embedded fast_float.
#include <cfloat>
#include <cstdint>

namespace tinyobj_ff {

// --- integral_constant, true_type, false_type ---
template <typename T, T V>
struct integral_constant {
  static const T value = V;
  typedef T value_type;
  typedef integral_constant type;
  operator value_type() const { return value; }
};
typedef integral_constant<bool, true>  true_type;
typedef integral_constant<bool, false> false_type;

// --- is_same ---
template <typename T, typename U> struct is_same       : false_type {};
template <typename T>             struct is_same<T, T> : true_type  {};

// --- enable_if ---
template <bool B, typename T = void> struct enable_if {};
template <typename T>                struct enable_if<true, T> { typedef T type; };

// --- conditional ---
template <bool B, typename T, typename F> struct conditional              { typedef T type; };
template <typename T, typename F>         struct conditional<false, T, F> { typedef F type; };

// --- is_integral ---
template <typename T> struct is_integral : false_type {};
template <> struct is_integral<bool>               : true_type {};
template <> struct is_integral<char>               : true_type {};
template <> struct is_integral<signed char>        : true_type {};
template <> struct is_integral<unsigned char>      : true_type {};
template <> struct is_integral<short>              : true_type {};
template <> struct is_integral<unsigned short>     : true_type {};
template <> struct is_integral<int>                : true_type {};
template <> struct is_integral<unsigned int>       : true_type {};
template <> struct is_integral<long>               : true_type {};
template <> struct is_integral<unsigned long>      : true_type {};
template <> struct is_integral<long long>          : true_type {};
template <> struct is_integral<unsigned long long> : true_type {};
template <> struct is_integral<wchar_t>            : true_type {};
template <> struct is_integral<char16_t>           : true_type {};
template <> struct is_integral<char32_t>           : true_type {};

// --- is_signed ---
template <typename T> struct is_signed : integral_constant<bool, T(-1) < T(0)> {};

// --- underlying_type (uses compiler builtin) ---
template <typename T> struct underlying_type {
  typedef __underlying_type(T) type;
};

// --- ff_errc (replaces std::errc, our own enum - no system_error needed) ---
enum class ff_errc { ok = 0, invalid_argument = 22, result_out_of_range = 34 };

// --- min_val (replaces std::min, avoids Windows min/max macro conflicts) ---
template <typename T>
inline T min_val(T a, T b) { return (b < a) ? b : a; }

// --- copy_n ---
template <typename InputIt, typename Size, typename OutputIt>
inline OutputIt copy_n(InputIt first, Size count, OutputIt result) {
  for (Size i = 0; i < count; ++i) *result++ = *first++;
  return result;
}

// --- copy_backward ---
template <typename BidirIt1, typename BidirIt2>
inline BidirIt2 copy_backward(BidirIt1 first, BidirIt1 last, BidirIt2 d_last) {
  while (first != last) *(--d_last) = *(--last);
  return d_last;
}

// --- fill ---
template <typename ForwardIt, typename T>
inline void fill(ForwardIt first, ForwardIt last, const T &value) {
  for (; first != last; ++first) *first = value;
}

// --- distance ---
template <typename It>
inline typename conditional<true, long long, It>::type
distance(It first, It last) {
  return last - first;
}

}  // namespace tinyobj_ff

// --- Begin embedded fast_float v8.0.2 (MIT / Apache-2.0 / BSL-1.0) ---
// https://github.com/fastfloat/fast_float
// fast_float by Daniel Lemire
// fast_float by João Paulo Magalhaes
//
//
// with contributions from Eugene Golushkov
// with contributions from Maksim Kita
// with contributions from Marcin Wojdyr
// with contributions from Neal Richardson
// with contributions from Tim Paine
// with contributions from Fabio Pellacini
// with contributions from Lénárd Szolnoki
// with contributions from Jan Pharago
// with contributions from Maya Warrier
// with contributions from Taha Khokhar
// with contributions from Anders Dalvander
//
//
// Licensed under the Apache License, Version 2.0, or the
// MIT License or the Boost License. This file may not be copied,
// modified, or distributed except according to those terms.
//
// MIT License Notice
//
//    MIT License
//
//    Copyright (c) 2021 The fast_float authors
//
//    Permission is hereby granted, free of charge, to any
//    person obtaining a copy of this software and associated
//    documentation files (the "Software"), to deal in the
//    Software without restriction, including without
//    limitation the rights to use, copy, modify, merge,
//    publish, distribute, sublicense, and/or sell copies of
//    the Software, and to permit persons to whom the Software
//    is furnished to do so, subject to the following
//    conditions:
//
//    The above copyright notice and this permission notice
//    shall be included in all copies or substantial portions
//    of the Software.
//
//    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF
//    ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED
//    TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A
//    PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT
//    SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
//    CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
//    OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR
//    IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
//    DEALINGS IN THE SOFTWARE.
//
// Apache License (Version 2.0) Notice
//
//    Copyright 2021 The fast_float authors
//    Licensed under the Apache License, Version 2.0 (the "License");
//    you may not use this file except in compliance with the License.
//    You may obtain a copy of the License at
//
//    http://www.apache.org/licenses/LICENSE-2.0
//
//    Unless required by applicable law or agreed to in writing, software
//    distributed under the License is distributed on an "AS IS" BASIS,
//    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
//    See the License for the specific language governing permissions and
//
// BOOST License Notice
//
//    Boost Software License - Version 1.0 - August 17th, 2003
//
//    Permission is hereby granted, free of charge, to any person or organization
//    obtaining a copy of the software and accompanying documentation covered by
//    this license (the "Software") to use, reproduce, display, distribute,
//    execute, and transmit the Software, and to prepare derivative works of the
//    Software, and to permit third-parties to whom the Software is furnished to
//    do so, all subject to the following:
//
//    The copyright notices in the Software and this entire statement, including
//    the above license grant, this restriction and the following disclaimer,
//    must be included in all copies of the Software, in whole or in part, and
//    all derivative works of the Software, unless such copies or derivative
//    works are solely in the form of machine-executable object code generated by
//    a source language processor.
//
//    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
//    IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
//    FITNESS FOR A PARTICULAR PURPOSE, TITLE AND NON-INFRINGEMENT. IN NO EVENT
//    SHALL THE COPYRIGHT HOLDERS OR ANYONE DISTRIBUTING THE SOFTWARE BE LIABLE
//    FOR ANY DAMAGES OR OTHER LIABILITY, WHETHER IN CONTRACT, TORT OR OTHERWISE,
//    ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
//    DEALINGS IN THE SOFTWARE.
//

#ifndef FASTFLOAT_CONSTEXPR_FEATURE_DETECT_H
#define FASTFLOAT_CONSTEXPR_FEATURE_DETECT_H

#ifdef __has_include
#if __has_include(<version>)
#include <version>
#endif
#endif

// Testing for https://wg21.link/N3652, adopted in C++14
#if defined(__cpp_constexpr) && __cpp_constexpr >= 201304
#define FASTFLOAT_CONSTEXPR14 constexpr
#else
#define FASTFLOAT_CONSTEXPR14
#endif

#if defined(__cpp_lib_bit_cast) && __cpp_lib_bit_cast >= 201806L
#define FASTFLOAT_HAS_BIT_CAST 1
#else
#define FASTFLOAT_HAS_BIT_CAST 0
#endif

#if defined(__cpp_lib_is_constant_evaluated) &&                                \
    __cpp_lib_is_constant_evaluated >= 201811L
#define FASTFLOAT_HAS_IS_CONSTANT_EVALUATED 1
#else
#define FASTFLOAT_HAS_IS_CONSTANT_EVALUATED 0
#endif

#if defined(__cpp_if_constexpr) && __cpp_if_constexpr >= 201606L
#define FASTFLOAT_IF_CONSTEXPR17(x) if constexpr (x)
#else
#define FASTFLOAT_IF_CONSTEXPR17(x) if (x)
#endif

// Testing for relevant C++20 constexpr library features
#if FASTFLOAT_HAS_IS_CONSTANT_EVALUATED && FASTFLOAT_HAS_BIT_CAST &&           \
    defined(__cpp_lib_constexpr_algorithms) &&                                 \
    __cpp_lib_constexpr_algorithms >= 201806L /*For std::copy and std::fill*/
#define FASTFLOAT_CONSTEXPR20 constexpr
#define FASTFLOAT_IS_CONSTEXPR 1
#else
#define FASTFLOAT_CONSTEXPR20
#define FASTFLOAT_IS_CONSTEXPR 0
#endif

#if __cplusplus >= 201703L || (defined(_MSVC_LANG) && _MSVC_LANG >= 201703L)
#define FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE 0
#else
#define FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE 1
#endif

#endif // FASTFLOAT_CONSTEXPR_FEATURE_DETECT_H

#ifndef FASTFLOAT_FLOAT_COMMON_H
#define FASTFLOAT_FLOAT_COMMON_H

#include <cassert>
#include <cstring>
#include <limits>
#ifdef __has_include
#if __has_include(<stdfloat>) && (__cplusplus > 202002L || (defined(_MSVC_LANG) && (_MSVC_LANG > 202002L)))
#include <stdfloat>
#endif
#endif

#define FASTFLOAT_VERSION_MAJOR 8
#define FASTFLOAT_VERSION_MINOR 0
#define FASTFLOAT_VERSION_PATCH 2

#define FASTFLOAT_STRINGIZE_IMPL(x) #x
#define FASTFLOAT_STRINGIZE(x) FASTFLOAT_STRINGIZE_IMPL(x)

#define FASTFLOAT_VERSION_STR                                                  \
  FASTFLOAT_STRINGIZE(FASTFLOAT_VERSION_MAJOR)                                 \
  "." FASTFLOAT_STRINGIZE(FASTFLOAT_VERSION_MINOR) "." FASTFLOAT_STRINGIZE(    \
      FASTFLOAT_VERSION_PATCH)

#define FASTFLOAT_VERSION                                                      \
  (FASTFLOAT_VERSION_MAJOR * 10000 + FASTFLOAT_VERSION_MINOR * 100 +           \
   FASTFLOAT_VERSION_PATCH)

namespace fast_float {

enum class chars_format : uint64_t;

namespace detail {
constexpr chars_format basic_json_fmt = chars_format(1 << 5);
constexpr chars_format basic_fortran_fmt = chars_format(1 << 6);
} // namespace detail

enum class chars_format : uint64_t {
  scientific = 1 << 0,
  fixed = 1 << 2,
  hex = 1 << 3,
  no_infnan = 1 << 4,
  // RFC 8259: https://datatracker.ietf.org/doc/html/rfc8259#section-6
  json = uint64_t(detail::basic_json_fmt) | fixed | scientific | no_infnan,
  // Extension of RFC 8259 where, e.g., "inf" and "nan" are allowed.
  json_or_infnan = uint64_t(detail::basic_json_fmt) | fixed | scientific,
  fortran = uint64_t(detail::basic_fortran_fmt) | fixed | scientific,
  general = fixed | scientific,
  allow_leading_plus = 1 << 7,
  skip_white_space = 1 << 8,
};

template <typename UC> struct from_chars_result_t {
  UC const *ptr;
  tinyobj_ff::ff_errc ec;
};

using from_chars_result = from_chars_result_t<char>;

template <typename UC> struct parse_options_t {
  constexpr explicit parse_options_t(chars_format fmt = chars_format::general,
                                     UC dot = UC('.'), int b = 10)
      : format(fmt), decimal_point(dot), base(b) {}

  /** Which number formats are accepted */
  chars_format format;
  /** The character used as decimal point */
  UC decimal_point;
  /** The base used for integers */
  int base;
};

using parse_options = parse_options_t<char>;

} // namespace fast_float

#if FASTFLOAT_HAS_BIT_CAST
#include <bit>
#endif

#if (defined(__x86_64) || defined(__x86_64__) || defined(_M_X64) ||            \
     defined(__amd64) || defined(__aarch64__) || defined(_M_ARM64) ||          \
     defined(__MINGW64__) || defined(__s390x__) ||                             \
     (defined(__ppc64__) || defined(__PPC64__) || defined(__ppc64le__) ||      \
      defined(__PPC64LE__)) ||                                                 \
     defined(__loongarch64))
#define FASTFLOAT_64BIT 1
#elif (defined(__i386) || defined(__i386__) || defined(_M_IX86) ||             \
       defined(__arm__) || defined(_M_ARM) || defined(__ppc__) ||              \
       defined(__MINGW32__) || defined(__EMSCRIPTEN__))
#define FASTFLOAT_32BIT 1
#else
  // Need to check incrementally, since SIZE_MAX is a size_t, avoid overflow.
// We can never tell the register width, but the SIZE_MAX is a good
// approximation. UINTPTR_MAX and INTPTR_MAX are optional, so avoid them for max
// portability.
#if SIZE_MAX == 0xffff
#error Unknown platform (16-bit, unsupported)
#elif SIZE_MAX == 0xffffffff
#define FASTFLOAT_32BIT 1
#elif SIZE_MAX == 0xffffffffffffffff
#define FASTFLOAT_64BIT 1
#else
#error Unknown platform (not 32-bit, not 64-bit?)
#endif
#endif

#if ((defined(_WIN32) || defined(_WIN64)) && !defined(__clang__)) ||           \
    (defined(_M_ARM64) && !defined(__MINGW32__))
#include <intrin.h>
#endif

#if defined(_MSC_VER) && !defined(__clang__)
#define FASTFLOAT_VISUAL_STUDIO 1
#endif

#if defined __BYTE_ORDER__ && defined __ORDER_BIG_ENDIAN__
#define FASTFLOAT_IS_BIG_ENDIAN (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
#elif defined _WIN32
#define FASTFLOAT_IS_BIG_ENDIAN 0
#else
#if defined(__APPLE__) || defined(__FreeBSD__)
#include <machine/endian.h>
#elif defined(sun) || defined(__sun)
#include <sys/byteorder.h>
#elif defined(__MVS__)
#include <sys/endian.h>
#else
#ifdef __has_include
#if __has_include(<endian.h>)
#include <endian.h>
#endif //__has_include(<endian.h>)
#endif //__has_include
#endif
#
#ifndef __BYTE_ORDER__
// safe choice
#define FASTFLOAT_IS_BIG_ENDIAN 0
#endif
#
#ifndef __ORDER_LITTLE_ENDIAN__
// safe choice
#define FASTFLOAT_IS_BIG_ENDIAN 0
#endif
#
#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
#define FASTFLOAT_IS_BIG_ENDIAN 0
#else
#define FASTFLOAT_IS_BIG_ENDIAN 1
#endif
#endif

#if defined(__SSE2__) || (defined(FASTFLOAT_VISUAL_STUDIO) &&                  \
                          (defined(_M_AMD64) || defined(_M_X64) ||             \
                           (defined(_M_IX86_FP) && _M_IX86_FP == 2)))
#define FASTFLOAT_SSE2 1
#endif

#if defined(__aarch64__) || defined(_M_ARM64)
#define FASTFLOAT_NEON 1
#endif

#if defined(FASTFLOAT_SSE2) || defined(FASTFLOAT_NEON)
#define FASTFLOAT_HAS_SIMD 1
#endif

#if defined(__GNUC__)
// disable -Wcast-align=strict (GCC only)
#define FASTFLOAT_SIMD_DISABLE_WARNINGS                                        \
  _Pragma("GCC diagnostic push")                                               \
      _Pragma("GCC diagnostic ignored \"-Wcast-align\"")
#else
#define FASTFLOAT_SIMD_DISABLE_WARNINGS
#endif

#if defined(__GNUC__)
#define FASTFLOAT_SIMD_RESTORE_WARNINGS _Pragma("GCC diagnostic pop")
#else
#define FASTFLOAT_SIMD_RESTORE_WARNINGS
#endif

#ifdef FASTFLOAT_VISUAL_STUDIO
#define fastfloat_really_inline __forceinline
#else
#define fastfloat_really_inline inline __attribute__((always_inline))
#endif

#ifndef FASTFLOAT_ASSERT
#define FASTFLOAT_ASSERT(x)                                                    \
  { ((void)(x)); }
#endif

#ifndef FASTFLOAT_DEBUG_ASSERT
#define FASTFLOAT_DEBUG_ASSERT(x)                                              \
  { ((void)(x)); }
#endif

// rust style `try!()` macro, or `?` operator
#define FASTFLOAT_TRY(x)                                                       \
  {                                                                            \
    if (!(x))                                                                  \
      return false;                                                            \
  }

#define FASTFLOAT_ENABLE_IF(...)                                               \
  typename tinyobj_ff::enable_if<(__VA_ARGS__), int>::type

namespace fast_float {

fastfloat_really_inline constexpr bool cpp20_and_in_constexpr() {
#if FASTFLOAT_HAS_IS_CONSTANT_EVALUATED
  return std::is_constant_evaluated();
#else
  return false;
#endif
}

template <typename T>
struct is_supported_float_type
    : tinyobj_ff::integral_constant<
          bool, tinyobj_ff::is_same<T, double>::value || tinyobj_ff::is_same<T, float>::value
#ifdef __STDCPP_FLOAT64_T__
                    || tinyobj_ff::is_same<T, std::float64_t>::value
#endif
#ifdef __STDCPP_FLOAT32_T__
                    || tinyobj_ff::is_same<T, std::float32_t>::value
#endif
#ifdef __STDCPP_FLOAT16_T__
                    || tinyobj_ff::is_same<T, std::float16_t>::value
#endif
#ifdef __STDCPP_BFLOAT16_T__
                    || tinyobj_ff::is_same<T, std::bfloat16_t>::value
#endif
          > {
};

template <typename T>
using equiv_uint_t = typename tinyobj_ff::conditional<
    sizeof(T) == 1, uint8_t,
    typename tinyobj_ff::conditional<
        sizeof(T) == 2, uint16_t,
        typename tinyobj_ff::conditional<sizeof(T) == 4, uint32_t,
                                  uint64_t>::type>::type>::type;

template <typename T> struct is_supported_integer_type : tinyobj_ff::is_integral<T> {};

template <typename UC>
struct is_supported_char_type
    : tinyobj_ff::integral_constant<bool, tinyobj_ff::is_same<UC, char>::value ||
                                       tinyobj_ff::is_same<UC, wchar_t>::value ||
                                       tinyobj_ff::is_same<UC, char16_t>::value ||
                                       tinyobj_ff::is_same<UC, char32_t>::value
#ifdef __cpp_char8_t
                                       || tinyobj_ff::is_same<UC, char8_t>::value
#endif
                             > {
};

// Compares two ASCII strings in a case insensitive manner.
template <typename UC>
inline FASTFLOAT_CONSTEXPR14 bool
fastfloat_strncasecmp(UC const *actual_mixedcase, UC const *expected_lowercase,
                      size_t length) {
  for (size_t i = 0; i < length; ++i) {
    UC const actual = actual_mixedcase[i];
    if ((actual < 256 ? actual | 32 : actual) != expected_lowercase[i]) {
      return false;
    }
  }
  return true;
}

#ifndef FLT_EVAL_METHOD
#error "FLT_EVAL_METHOD should be defined, please include cfloat."
#endif

// a pointer and a length to a contiguous block of memory
template <typename T> struct span {
  T const *ptr;
  size_t length;

  constexpr span(T const *_ptr, size_t _length) : ptr(_ptr), length(_length) {}

  constexpr span() : ptr(nullptr), length(0) {}

  constexpr size_t len() const noexcept { return length; }

  FASTFLOAT_CONSTEXPR14 const T &operator[](size_t index) const noexcept {
    FASTFLOAT_DEBUG_ASSERT(index < length);
    return ptr[index];
  }
};

struct value128 {
  uint64_t low;
  uint64_t high;

  constexpr value128(uint64_t _low, uint64_t _high) : low(_low), high(_high) {}

  constexpr value128() : low(0), high(0) {}
};

/* Helper C++14 constexpr generic implementation of leading_zeroes */
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 int
leading_zeroes_generic(uint64_t input_num, int last_bit = 0) {
  if (input_num & uint64_t(0xffffffff00000000)) {
    input_num >>= 32;
    last_bit |= 32;
  }
  if (input_num & uint64_t(0xffff0000)) {
    input_num >>= 16;
    last_bit |= 16;
  }
  if (input_num & uint64_t(0xff00)) {
    input_num >>= 8;
    last_bit |= 8;
  }
  if (input_num & uint64_t(0xf0)) {
    input_num >>= 4;
    last_bit |= 4;
  }
  if (input_num & uint64_t(0xc)) {
    input_num >>= 2;
    last_bit |= 2;
  }
  if (input_num & uint64_t(0x2)) { /* input_num >>=  1; */
    last_bit |= 1;
  }
  return 63 - last_bit;
}

/* result might be undefined when input_num is zero */
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 int
leading_zeroes(uint64_t input_num) {
  assert(input_num > 0);
  if (cpp20_and_in_constexpr()) {
    return leading_zeroes_generic(input_num);
  }
#ifdef FASTFLOAT_VISUAL_STUDIO
#if defined(_M_X64) || defined(_M_ARM64)
  unsigned long leading_zero = 0;
  // Search the mask data from most significant bit (MSB)
  // to least significant bit (LSB) for a set bit (1).
  _BitScanReverse64(&leading_zero, input_num);
  return (int)(63 - leading_zero);
#else
  return leading_zeroes_generic(input_num);
#endif
#else
  return __builtin_clzll(input_num);
#endif
}

// slow emulation routine for 32-bit
fastfloat_really_inline constexpr uint64_t emulu(uint32_t x, uint32_t y) {
  return x * (uint64_t)y;
}

fastfloat_really_inline FASTFLOAT_CONSTEXPR14 uint64_t
umul128_generic(uint64_t ab, uint64_t cd, uint64_t *hi) {
  uint64_t ad = emulu((uint32_t)(ab >> 32), (uint32_t)cd);
  uint64_t bd = emulu((uint32_t)ab, (uint32_t)cd);
  uint64_t adbc = ad + emulu((uint32_t)ab, (uint32_t)(cd >> 32));
  uint64_t adbc_carry = (uint64_t)(adbc < ad);
  uint64_t lo = bd + (adbc << 32);
  *hi = emulu((uint32_t)(ab >> 32), (uint32_t)(cd >> 32)) + (adbc >> 32) +
        (adbc_carry << 32) + (uint64_t)(lo < bd);
  return lo;
}

#ifdef FASTFLOAT_32BIT

// slow emulation routine for 32-bit
#if !defined(__MINGW64__)
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 uint64_t _umul128(uint64_t ab,
                                                                uint64_t cd,
                                                                uint64_t *hi) {
  return umul128_generic(ab, cd, hi);
}
#endif // !__MINGW64__

#endif // FASTFLOAT_32BIT

// compute 64-bit a*b
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 value128
full_multiplication(uint64_t a, uint64_t b) {
  if (cpp20_and_in_constexpr()) {
    value128 answer;
    answer.low = umul128_generic(a, b, &answer.high);
    return answer;
  }
  value128 answer;
#if defined(_M_ARM64) && !defined(__MINGW32__)
  // ARM64 has native support for 64-bit multiplications, no need to emulate
  // But MinGW on ARM64 doesn't have native support for 64-bit multiplications
  answer.high = __umulh(a, b);
  answer.low = a * b;
#elif defined(FASTFLOAT_32BIT) ||                                              \
    (defined(_WIN64) && !defined(__clang__) && !defined(_M_ARM64))
  answer.low = _umul128(a, b, &answer.high); // _umul128 not available on ARM64
#elif defined(FASTFLOAT_64BIT) && defined(__SIZEOF_INT128__)
  __uint128_t r = ((__uint128_t)a) * b;
  answer.low = uint64_t(r);
  answer.high = uint64_t(r >> 64);
#else
  answer.low = umul128_generic(a, b, &answer.high);
#endif
  return answer;
}

struct adjusted_mantissa {
  uint64_t mantissa{0};
  int32_t power2{0}; // a negative value indicates an invalid result
  adjusted_mantissa() = default;

  constexpr bool operator==(adjusted_mantissa const &o) const {
    return mantissa == o.mantissa && power2 == o.power2;
  }

  constexpr bool operator!=(adjusted_mantissa const &o) const {
    return mantissa != o.mantissa || power2 != o.power2;
  }
};

// Bias so we can get the real exponent with an invalid adjusted_mantissa.
constexpr static int32_t invalid_am_bias = -0x8000;

// used for binary_format_lookup_tables<T>::max_mantissa
constexpr uint64_t constant_55555 = 5 * 5 * 5 * 5 * 5;

template <typename T, typename U = void> struct binary_format_lookup_tables;

template <typename T> struct binary_format : binary_format_lookup_tables<T> {
  using equiv_uint = equiv_uint_t<T>;

  static constexpr int mantissa_explicit_bits();
  static constexpr int minimum_exponent();
  static constexpr int infinite_power();
  static constexpr int sign_index();
  static constexpr int
  min_exponent_fast_path(); // used when fegetround() == FE_TONEAREST
  static constexpr int max_exponent_fast_path();
  static constexpr int max_exponent_round_to_even();
  static constexpr int min_exponent_round_to_even();
  static constexpr uint64_t max_mantissa_fast_path(int64_t power);
  static constexpr uint64_t
  max_mantissa_fast_path(); // used when fegetround() == FE_TONEAREST
  static constexpr int largest_power_of_ten();
  static constexpr int smallest_power_of_ten();
  static constexpr T exact_power_of_ten(int64_t power);
  static constexpr size_t max_digits();
  static constexpr equiv_uint exponent_mask();
  static constexpr equiv_uint mantissa_mask();
  static constexpr equiv_uint hidden_bit_mask();
};

template <typename U> struct binary_format_lookup_tables<double, U> {
  static constexpr double powers_of_ten[] = {
      1e0,  1e1,  1e2,  1e3,  1e4,  1e5,  1e6,  1e7,  1e8,  1e9,  1e10, 1e11,
      1e12, 1e13, 1e14, 1e15, 1e16, 1e17, 1e18, 1e19, 1e20, 1e21, 1e22};

  // Largest integer value v so that (5**index * v) <= 1<<53.
  // 0x20000000000000 == 1 << 53
  static constexpr uint64_t max_mantissa[] = {
      0x20000000000000,
      0x20000000000000 / 5,
      0x20000000000000 / (5 * 5),
      0x20000000000000 / (5 * 5 * 5),
      0x20000000000000 / (5 * 5 * 5 * 5),
      0x20000000000000 / (constant_55555),
      0x20000000000000 / (constant_55555 * 5),
      0x20000000000000 / (constant_55555 * 5 * 5),
      0x20000000000000 / (constant_55555 * 5 * 5 * 5),
      0x20000000000000 / (constant_55555 * 5 * 5 * 5 * 5),
      0x20000000000000 / (constant_55555 * constant_55555),
      0x20000000000000 / (constant_55555 * constant_55555 * 5),
      0x20000000000000 / (constant_55555 * constant_55555 * 5 * 5),
      0x20000000000000 / (constant_55555 * constant_55555 * 5 * 5 * 5),
      0x20000000000000 / (constant_55555 * constant_55555 * constant_55555),
      0x20000000000000 / (constant_55555 * constant_55555 * constant_55555 * 5),
      0x20000000000000 /
          (constant_55555 * constant_55555 * constant_55555 * 5 * 5),
      0x20000000000000 /
          (constant_55555 * constant_55555 * constant_55555 * 5 * 5 * 5),
      0x20000000000000 /
          (constant_55555 * constant_55555 * constant_55555 * 5 * 5 * 5 * 5),
      0x20000000000000 /
          (constant_55555 * constant_55555 * constant_55555 * constant_55555),
      0x20000000000000 / (constant_55555 * constant_55555 * constant_55555 *
                          constant_55555 * 5),
      0x20000000000000 / (constant_55555 * constant_55555 * constant_55555 *
                          constant_55555 * 5 * 5),
      0x20000000000000 / (constant_55555 * constant_55555 * constant_55555 *
                          constant_55555 * 5 * 5 * 5),
      0x20000000000000 / (constant_55555 * constant_55555 * constant_55555 *
                          constant_55555 * 5 * 5 * 5 * 5)};
};

#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE

template <typename U>
constexpr double binary_format_lookup_tables<double, U>::powers_of_ten[];

template <typename U>
constexpr uint64_t binary_format_lookup_tables<double, U>::max_mantissa[];

#endif

template <typename U> struct binary_format_lookup_tables<float, U> {
  static constexpr float powers_of_ten[] = {1e0f, 1e1f, 1e2f, 1e3f, 1e4f, 1e5f,
                                            1e6f, 1e7f, 1e8f, 1e9f, 1e10f};

  // Largest integer value v so that (5**index * v) <= 1<<24.
  // 0x1000000 == 1<<24
  static constexpr uint64_t max_mantissa[] = {
      0x1000000,
      0x1000000 / 5,
      0x1000000 / (5 * 5),
      0x1000000 / (5 * 5 * 5),
      0x1000000 / (5 * 5 * 5 * 5),
      0x1000000 / (constant_55555),
      0x1000000 / (constant_55555 * 5),
      0x1000000 / (constant_55555 * 5 * 5),
      0x1000000 / (constant_55555 * 5 * 5 * 5),
      0x1000000 / (constant_55555 * 5 * 5 * 5 * 5),
      0x1000000 / (constant_55555 * constant_55555),
      0x1000000 / (constant_55555 * constant_55555 * 5)};
};

#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE

template <typename U>
constexpr float binary_format_lookup_tables<float, U>::powers_of_ten[];

template <typename U>
constexpr uint64_t binary_format_lookup_tables<float, U>::max_mantissa[];

#endif

template <>
inline constexpr int binary_format<double>::min_exponent_fast_path() {
#if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0)
  return 0;
#else
  return -22;
#endif
}

template <>
inline constexpr int binary_format<float>::min_exponent_fast_path() {
#if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0)
  return 0;
#else
  return -10;
#endif
}

template <>
inline constexpr int binary_format<double>::mantissa_explicit_bits() {
  return 52;
}

template <>
inline constexpr int binary_format<float>::mantissa_explicit_bits() {
  return 23;
}

template <>
inline constexpr int binary_format<double>::max_exponent_round_to_even() {
  return 23;
}

template <>
inline constexpr int binary_format<float>::max_exponent_round_to_even() {
  return 10;
}

template <>
inline constexpr int binary_format<double>::min_exponent_round_to_even() {
  return -4;
}

template <>
inline constexpr int binary_format<float>::min_exponent_round_to_even() {
  return -17;
}

template <> inline constexpr int binary_format<double>::minimum_exponent() {
  return -1023;
}

template <> inline constexpr int binary_format<float>::minimum_exponent() {
  return -127;
}

template <> inline constexpr int binary_format<double>::infinite_power() {
  return 0x7FF;
}

template <> inline constexpr int binary_format<float>::infinite_power() {
  return 0xFF;
}

template <> inline constexpr int binary_format<double>::sign_index() {
  return 63;
}

template <> inline constexpr int binary_format<float>::sign_index() {
  return 31;
}

template <>
inline constexpr int binary_format<double>::max_exponent_fast_path() {
  return 22;
}

template <>
inline constexpr int binary_format<float>::max_exponent_fast_path() {
  return 10;
}

template <>
inline constexpr uint64_t binary_format<double>::max_mantissa_fast_path() {
  return uint64_t(2) << mantissa_explicit_bits();
}

template <>
inline constexpr uint64_t binary_format<float>::max_mantissa_fast_path() {
  return uint64_t(2) << mantissa_explicit_bits();
}

// credit: Jakub Jelínek
#ifdef __STDCPP_FLOAT16_T__
template <typename U> struct binary_format_lookup_tables<std::float16_t, U> {
  static constexpr std::float16_t powers_of_ten[] = {1e0f16, 1e1f16, 1e2f16,
                                                     1e3f16, 1e4f16};

  // Largest integer value v so that (5**index * v) <= 1<<11.
  // 0x800 == 1<<11
  static constexpr uint64_t max_mantissa[] = {0x800,
                                              0x800 / 5,
                                              0x800 / (5 * 5),
                                              0x800 / (5 * 5 * 5),
                                              0x800 / (5 * 5 * 5 * 5),
                                              0x800 / (constant_55555)};
};

#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE

template <typename U>
constexpr std::float16_t
    binary_format_lookup_tables<std::float16_t, U>::powers_of_ten[];

template <typename U>
constexpr uint64_t
    binary_format_lookup_tables<std::float16_t, U>::max_mantissa[];

#endif

template <>
inline constexpr std::float16_t
binary_format<std::float16_t>::exact_power_of_ten(int64_t power) {
  // Work around clang bug https://godbolt.org/z/zedh7rrhc
  return (void)powers_of_ten[0], powers_of_ten[power];
}

template <>
inline constexpr binary_format<std::float16_t>::equiv_uint
binary_format<std::float16_t>::exponent_mask() {
  return 0x7C00;
}

template <>
inline constexpr binary_format<std::float16_t>::equiv_uint
binary_format<std::float16_t>::mantissa_mask() {
  return 0x03FF;
}

template <>
inline constexpr binary_format<std::float16_t>::equiv_uint
binary_format<std::float16_t>::hidden_bit_mask() {
  return 0x0400;
}

template <>
inline constexpr int binary_format<std::float16_t>::max_exponent_fast_path() {
  return 4;
}

template <>
inline constexpr int binary_format<std::float16_t>::mantissa_explicit_bits() {
  return 10;
}

template <>
inline constexpr uint64_t
binary_format<std::float16_t>::max_mantissa_fast_path() {
  return uint64_t(2) << mantissa_explicit_bits();
}

template <>
inline constexpr uint64_t
binary_format<std::float16_t>::max_mantissa_fast_path(int64_t power) {
  // caller is responsible to ensure that
  // power >= 0 && power <= 4
  //
  // Work around clang bug https://godbolt.org/z/zedh7rrhc
  return (void)max_mantissa[0], max_mantissa[power];
}

template <>
inline constexpr int binary_format<std::float16_t>::min_exponent_fast_path() {
  return 0;
}

template <>
inline constexpr int
binary_format<std::float16_t>::max_exponent_round_to_even() {
  return 5;
}

template <>
inline constexpr int
binary_format<std::float16_t>::min_exponent_round_to_even() {
  return -22;
}

template <>
inline constexpr int binary_format<std::float16_t>::minimum_exponent() {
  return -15;
}

template <>
inline constexpr int binary_format<std::float16_t>::infinite_power() {
  return 0x1F;
}

template <> inline constexpr int binary_format<std::float16_t>::sign_index() {
  return 15;
}

template <>
inline constexpr int binary_format<std::float16_t>::largest_power_of_ten() {
  return 4;
}

template <>
inline constexpr int binary_format<std::float16_t>::smallest_power_of_ten() {
  return -27;
}

template <>
inline constexpr size_t binary_format<std::float16_t>::max_digits() {
  return 22;
}
#endif // __STDCPP_FLOAT16_T__

// credit: Jakub Jelínek
#ifdef __STDCPP_BFLOAT16_T__
template <typename U> struct binary_format_lookup_tables<std::bfloat16_t, U> {
  static constexpr std::bfloat16_t powers_of_ten[] = {1e0bf16, 1e1bf16, 1e2bf16,
                                                      1e3bf16};

  // Largest integer value v so that (5**index * v) <= 1<<8.
  // 0x100 == 1<<8
  static constexpr uint64_t max_mantissa[] = {0x100, 0x100 / 5, 0x100 / (5 * 5),
                                              0x100 / (5 * 5 * 5),
                                              0x100 / (5 * 5 * 5 * 5)};
};

#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE

template <typename U>
constexpr std::bfloat16_t
    binary_format_lookup_tables<std::bfloat16_t, U>::powers_of_ten[];

template <typename U>
constexpr uint64_t
    binary_format_lookup_tables<std::bfloat16_t, U>::max_mantissa[];

#endif

template <>
inline constexpr std::bfloat16_t
binary_format<std::bfloat16_t>::exact_power_of_ten(int64_t power) {
  // Work around clang bug https://godbolt.org/z/zedh7rrhc
  return (void)powers_of_ten[0], powers_of_ten[power];
}

template <>
inline constexpr int binary_format<std::bfloat16_t>::max_exponent_fast_path() {
  return 3;
}

template <>
inline constexpr binary_format<std::bfloat16_t>::equiv_uint
binary_format<std::bfloat16_t>::exponent_mask() {
  return 0x7F80;
}

template <>
inline constexpr binary_format<std::bfloat16_t>::equiv_uint
binary_format<std::bfloat16_t>::mantissa_mask() {
  return 0x007F;
}

template <>
inline constexpr binary_format<std::bfloat16_t>::equiv_uint
binary_format<std::bfloat16_t>::hidden_bit_mask() {
  return 0x0080;
}

template <>
inline constexpr int binary_format<std::bfloat16_t>::mantissa_explicit_bits() {
  return 7;
}

template <>
inline constexpr uint64_t
binary_format<std::bfloat16_t>::max_mantissa_fast_path() {
  return uint64_t(2) << mantissa_explicit_bits();
}

template <>
inline constexpr uint64_t
binary_format<std::bfloat16_t>::max_mantissa_fast_path(int64_t power) {
  // caller is responsible to ensure that
  // power >= 0 && power <= 3
  //
  // Work around clang bug https://godbolt.org/z/zedh7rrhc
  return (void)max_mantissa[0], max_mantissa[power];
}

template <>
inline constexpr int binary_format<std::bfloat16_t>::min_exponent_fast_path() {
  return 0;
}

template <>
inline constexpr int
binary_format<std::bfloat16_t>::max_exponent_round_to_even() {
  return 3;
}

template <>
inline constexpr int
binary_format<std::bfloat16_t>::min_exponent_round_to_even() {
  return -24;
}

template <>
inline constexpr int binary_format<std::bfloat16_t>::minimum_exponent() {
  return -127;
}

template <>
inline constexpr int binary_format<std::bfloat16_t>::infinite_power() {
  return 0xFF;
}

template <> inline constexpr int binary_format<std::bfloat16_t>::sign_index() {
  return 15;
}

template <>
inline constexpr int binary_format<std::bfloat16_t>::largest_power_of_ten() {
  return 38;
}

template <>
inline constexpr int binary_format<std::bfloat16_t>::smallest_power_of_ten() {
  return -60;
}

template <>
inline constexpr size_t binary_format<std::bfloat16_t>::max_digits() {
  return 98;
}
#endif // __STDCPP_BFLOAT16_T__

template <>
inline constexpr uint64_t
binary_format<double>::max_mantissa_fast_path(int64_t power) {
  // caller is responsible to ensure that
  // power >= 0 && power <= 22
  //
  // Work around clang bug https://godbolt.org/z/zedh7rrhc
  return (void)max_mantissa[0], max_mantissa[power];
}

template <>
inline constexpr uint64_t
binary_format<float>::max_mantissa_fast_path(int64_t power) {
  // caller is responsible to ensure that
  // power >= 0 && power <= 10
  //
  // Work around clang bug https://godbolt.org/z/zedh7rrhc
  return (void)max_mantissa[0], max_mantissa[power];
}

template <>
inline constexpr double
binary_format<double>::exact_power_of_ten(int64_t power) {
  // Work around clang bug https://godbolt.org/z/zedh7rrhc
  return (void)powers_of_ten[0], powers_of_ten[power];
}

template <>
inline constexpr float binary_format<float>::exact_power_of_ten(int64_t power) {
  // Work around clang bug https://godbolt.org/z/zedh7rrhc
  return (void)powers_of_ten[0], powers_of_ten[power];
}

template <> inline constexpr int binary_format<double>::largest_power_of_ten() {
  return 308;
}

template <> inline constexpr int binary_format<float>::largest_power_of_ten() {
  return 38;
}

template <>
inline constexpr int binary_format<double>::smallest_power_of_ten() {
  return -342;
}

template <> inline constexpr int binary_format<float>::smallest_power_of_ten() {
  return -64;
}

template <> inline constexpr size_t binary_format<double>::max_digits() {
  return 769;
}

template <> inline constexpr size_t binary_format<float>::max_digits() {
  return 114;
}

template <>
inline constexpr binary_format<float>::equiv_uint
binary_format<float>::exponent_mask() {
  return 0x7F800000;
}

template <>
inline constexpr binary_format<double>::equiv_uint
binary_format<double>::exponent_mask() {
  return 0x7FF0000000000000;
}

template <>
inline constexpr binary_format<float>::equiv_uint
binary_format<float>::mantissa_mask() {
  return 0x007FFFFF;
}

template <>
inline constexpr binary_format<double>::equiv_uint
binary_format<double>::mantissa_mask() {
  return 0x000FFFFFFFFFFFFF;
}

template <>
inline constexpr binary_format<float>::equiv_uint
binary_format<float>::hidden_bit_mask() {
  return 0x00800000;
}

template <>
inline constexpr binary_format<double>::equiv_uint
binary_format<double>::hidden_bit_mask() {
  return 0x0010000000000000;
}

template <typename T>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
to_float(bool negative, adjusted_mantissa am, T &value) {
  using equiv_uint = equiv_uint_t<T>;
  equiv_uint word = equiv_uint(am.mantissa);
  word = equiv_uint(word | equiv_uint(am.power2)
                               << binary_format<T>::mantissa_explicit_bits());
  word =
      equiv_uint(word | equiv_uint(negative) << binary_format<T>::sign_index());
#if FASTFLOAT_HAS_BIT_CAST
  value = std::bit_cast<T>(word);
#else
  ::memcpy(&value, &word, sizeof(T));
#endif
}

template <typename = void> struct space_lut {
  static constexpr bool value[] = {
      0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
      0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
};

#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE

template <typename T> constexpr bool space_lut<T>::value[];

#endif

template <typename UC> constexpr bool is_space(UC c) {
  return c < 256 && space_lut<>::value[uint8_t(c)];
}

template <typename UC> static constexpr uint64_t int_cmp_zeros() {
  static_assert((sizeof(UC) == 1) || (sizeof(UC) == 2) || (sizeof(UC) == 4),
                "Unsupported character size");
  return (sizeof(UC) == 1) ? 0x3030303030303030
         : (sizeof(UC) == 2)
             ? (uint64_t(UC('0')) << 48 | uint64_t(UC('0')) << 32 |
                uint64_t(UC('0')) << 16 | UC('0'))
             : (uint64_t(UC('0')) << 32 | UC('0'));
}

template <typename UC> static constexpr int int_cmp_len() {
  return sizeof(uint64_t) / sizeof(UC);
}

template <typename UC> constexpr UC const *str_const_nan();

template <> constexpr char const *str_const_nan<char>() { return "nan"; }

template <> constexpr wchar_t const *str_const_nan<wchar_t>() { return L"nan"; }

template <> constexpr char16_t const *str_const_nan<char16_t>() {
  return u"nan";
}

template <> constexpr char32_t const *str_const_nan<char32_t>() {
  return U"nan";
}

#ifdef __cpp_char8_t
template <> constexpr char8_t const *str_const_nan<char8_t>() {
  return u8"nan";
}
#endif

template <typename UC> constexpr UC const *str_const_inf();

template <> constexpr char const *str_const_inf<char>() { return "infinity"; }

template <> constexpr wchar_t const *str_const_inf<wchar_t>() {
  return L"infinity";
}

template <> constexpr char16_t const *str_const_inf<char16_t>() {
  return u"infinity";
}

template <> constexpr char32_t const *str_const_inf<char32_t>() {
  return U"infinity";
}

#ifdef __cpp_char8_t
template <> constexpr char8_t const *str_const_inf<char8_t>() {
  return u8"infinity";
}
#endif

template <typename = void> struct int_luts {
  static constexpr uint8_t chdigit[] = {
      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
      255, 255, 255, 0,   1,   2,   3,   4,   5,   6,   7,   8,   9,   255, 255,
      255, 255, 255, 255, 255, 10,  11,  12,  13,  14,  15,  16,  17,  18,  19,
      20,  21,  22,  23,  24,  25,  26,  27,  28,  29,  30,  31,  32,  33,  34,
      35,  255, 255, 255, 255, 255, 255, 10,  11,  12,  13,  14,  15,  16,  17,
      18,  19,  20,  21,  22,  23,  24,  25,  26,  27,  28,  29,  30,  31,  32,
      33,  34,  35,  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
      255};

  static constexpr size_t maxdigits_u64[] = {
      64, 41, 32, 28, 25, 23, 22, 21, 20, 19, 18, 18, 17, 17, 16, 16, 16, 16,
      15, 15, 15, 15, 14, 14, 14, 14, 14, 14, 14, 13, 13, 13, 13, 13, 13};

  static constexpr uint64_t min_safe_u64[] = {
      9223372036854775808ull,  12157665459056928801ull, 4611686018427387904,
      7450580596923828125,     4738381338321616896,     3909821048582988049,
      9223372036854775808ull,  12157665459056928801ull, 10000000000000000000ull,
      5559917313492231481,     2218611106740436992,     8650415919381337933,
      2177953337809371136,     6568408355712890625,     1152921504606846976,
      2862423051509815793,     6746640616477458432,     15181127029874798299ull,
      1638400000000000000,     3243919932521508681,     6221821273427820544,
      11592836324538749809ull, 876488338465357824,      1490116119384765625,
      2481152873203736576,     4052555153018976267,     6502111422497947648,
      10260628712958602189ull, 15943230000000000000ull, 787662783788549761,
      1152921504606846976,     1667889514952984961,     2386420683693101056,
      3379220508056640625,     4738381338321616896};
};

#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE

template <typename T> constexpr uint8_t int_luts<T>::chdigit[];

template <typename T> constexpr size_t int_luts<T>::maxdigits_u64[];

template <typename T> constexpr uint64_t int_luts<T>::min_safe_u64[];

#endif

template <typename UC>
fastfloat_really_inline constexpr uint8_t ch_to_digit(UC c) {
  return int_luts<>::chdigit[static_cast<unsigned char>(c)];
}

fastfloat_really_inline constexpr size_t max_digits_u64(int base) {
  return int_luts<>::maxdigits_u64[base - 2];
}

// If a u64 is exactly max_digits_u64() in length, this is
// the value below which it has definitely overflowed.
fastfloat_really_inline constexpr uint64_t min_safe_u64(int base) {
  return int_luts<>::min_safe_u64[base - 2];
}

static_assert(tinyobj_ff::is_same<equiv_uint_t<double>, uint64_t>::value,
              "equiv_uint should be uint64_t for double");
static_assert(std::numeric_limits<double>::is_iec559,
              "double must fulfill the requirements of IEC 559 (IEEE 754)");

static_assert(tinyobj_ff::is_same<equiv_uint_t<float>, uint32_t>::value,
              "equiv_uint should be uint32_t for float");
static_assert(std::numeric_limits<float>::is_iec559,
              "float must fulfill the requirements of IEC 559 (IEEE 754)");

#ifdef __STDCPP_FLOAT64_T__
static_assert(tinyobj_ff::is_same<equiv_uint_t<std::float64_t>, uint64_t>::value,
              "equiv_uint should be uint64_t for std::float64_t");
static_assert(
    std::numeric_limits<std::float64_t>::is_iec559,
    "std::float64_t must fulfill the requirements of IEC 559 (IEEE 754)");
#endif // __STDCPP_FLOAT64_T__

#ifdef __STDCPP_FLOAT32_T__
static_assert(tinyobj_ff::is_same<equiv_uint_t<std::float32_t>, uint32_t>::value,
              "equiv_uint should be uint32_t for std::float32_t");
static_assert(
    std::numeric_limits<std::float32_t>::is_iec559,
    "std::float32_t must fulfill the requirements of IEC 559 (IEEE 754)");
#endif // __STDCPP_FLOAT32_T__

#ifdef __STDCPP_FLOAT16_T__
static_assert(
    tinyobj_ff::is_same<binary_format<std::float16_t>::equiv_uint, uint16_t>::value,
    "equiv_uint should be uint16_t for std::float16_t");
static_assert(
    std::numeric_limits<std::float16_t>::is_iec559,
    "std::float16_t must fulfill the requirements of IEC 559 (IEEE 754)");
#endif // __STDCPP_FLOAT16_T__

#ifdef __STDCPP_BFLOAT16_T__
static_assert(
    tinyobj_ff::is_same<binary_format<std::bfloat16_t>::equiv_uint, uint16_t>::value,
    "equiv_uint should be uint16_t for std::bfloat16_t");
static_assert(
    std::numeric_limits<std::bfloat16_t>::is_iec559,
    "std::bfloat16_t must fulfill the requirements of IEC 559 (IEEE 754)");
#endif // __STDCPP_BFLOAT16_T__

constexpr chars_format operator~(chars_format rhs) noexcept {
  using int_type = tinyobj_ff::underlying_type<chars_format>::type;
  return static_cast<chars_format>(~static_cast<int_type>(rhs));
}

constexpr chars_format operator&(chars_format lhs, chars_format rhs) noexcept {
  using int_type = tinyobj_ff::underlying_type<chars_format>::type;
  return static_cast<chars_format>(static_cast<int_type>(lhs) &
                                   static_cast<int_type>(rhs));
}

constexpr chars_format operator|(chars_format lhs, chars_format rhs) noexcept {
  using int_type = tinyobj_ff::underlying_type<chars_format>::type;
  return static_cast<chars_format>(static_cast<int_type>(lhs) |
                                   static_cast<int_type>(rhs));
}

constexpr chars_format operator^(chars_format lhs, chars_format rhs) noexcept {
  using int_type = tinyobj_ff::underlying_type<chars_format>::type;
  return static_cast<chars_format>(static_cast<int_type>(lhs) ^
                                   static_cast<int_type>(rhs));
}

fastfloat_really_inline FASTFLOAT_CONSTEXPR14 chars_format &
operator&=(chars_format &lhs, chars_format rhs) noexcept {
  return lhs = (lhs & rhs);
}

fastfloat_really_inline FASTFLOAT_CONSTEXPR14 chars_format &
operator|=(chars_format &lhs, chars_format rhs) noexcept {
  return lhs = (lhs | rhs);
}

fastfloat_really_inline FASTFLOAT_CONSTEXPR14 chars_format &
operator^=(chars_format &lhs, chars_format rhs) noexcept {
  return lhs = (lhs ^ rhs);
}

namespace detail {
// adjust for deprecated feature macros
constexpr chars_format adjust_for_feature_macros(chars_format fmt) {
  return fmt
#ifdef FASTFLOAT_ALLOWS_LEADING_PLUS
         | chars_format::allow_leading_plus
#endif
#ifdef FASTFLOAT_SKIP_WHITE_SPACE
         | chars_format::skip_white_space
#endif
      ;
}
} // namespace detail

} // namespace fast_float

#endif


#ifndef FASTFLOAT_FAST_FLOAT_H
#define FASTFLOAT_FAST_FLOAT_H


namespace fast_float {
/**
 * This function parses the character sequence [first,last) for a number. It
 * parses floating-point numbers expecting a locale-indepent format equivalent
 * to what is used by std::strtod in the default ("C") locale. The resulting
 * floating-point value is the closest floating-point values (using either float
 * or double), using the "round to even" convention for values that would
 * otherwise fall right in-between two values. That is, we provide exact parsing
 * according to the IEEE standard.
 *
 * Given a successful parse, the pointer (`ptr`) in the returned value is set to
 * point right after the parsed number, and the `value` referenced is set to the
 * parsed value. In case of error, the returned `ec` contains a representative
 * error, otherwise the default (`tinyobj_ff::ff_errc()`) value is stored.
 *
 * The implementation does not throw and does not allocate memory (e.g., with
 * `new` or `malloc`).
 *
 * Like the C++17 standard, the `fast_float::from_chars` functions take an
 * optional last argument of the type `fast_float::chars_format`. It is a bitset
 * value: we check whether `fmt & fast_float::chars_format::fixed` and `fmt &
 * fast_float::chars_format::scientific` are set to determine whether we allow
 * the fixed point and scientific notation respectively. The default is
 * `fast_float::chars_format::general` which allows both `fixed` and
 * `scientific`.
 */
template <typename T, typename UC = char,
          typename = FASTFLOAT_ENABLE_IF(is_supported_float_type<T>::value)>
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
from_chars(UC const *first, UC const *last, T &value,
           chars_format fmt = chars_format::general) noexcept;

/**
 * Like from_chars, but accepts an `options` argument to govern number parsing.
 * Both for floating-point types and integer types.
 */
template <typename T, typename UC = char>
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
from_chars_advanced(UC const *first, UC const *last, T &value,
                    parse_options_t<UC> options) noexcept;

/**
 * from_chars for integer types.
 */
template <typename T, typename UC = char,
          typename = FASTFLOAT_ENABLE_IF(is_supported_integer_type<T>::value)>
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
from_chars(UC const *first, UC const *last, T &value, int base = 10) noexcept;

} // namespace fast_float

#endif // FASTFLOAT_FAST_FLOAT_H

#ifndef FASTFLOAT_ASCII_NUMBER_H
#define FASTFLOAT_ASCII_NUMBER_H

#include <cctype>
#include <cstring>
#include <limits>


#ifdef FASTFLOAT_SSE2
#include <emmintrin.h>
#endif

#ifdef FASTFLOAT_NEON
#include <arm_neon.h>
#endif

namespace fast_float {

template <typename UC> fastfloat_really_inline constexpr bool has_simd_opt() {
#ifdef FASTFLOAT_HAS_SIMD
  return tinyobj_ff::is_same<UC, char16_t>::value;
#else
  return false;
#endif
}

// Next function can be micro-optimized, but compilers are entirely
// able to optimize it well.
template <typename UC>
fastfloat_really_inline constexpr bool is_integer(UC c) noexcept {
  return !(c > UC('9') || c < UC('0'));
}

fastfloat_really_inline constexpr uint64_t byteswap(uint64_t val) {
  return (val & 0xFF00000000000000) >> 56 | (val & 0x00FF000000000000) >> 40 |
         (val & 0x0000FF0000000000) >> 24 | (val & 0x000000FF00000000) >> 8 |
         (val & 0x00000000FF000000) << 8 | (val & 0x0000000000FF0000) << 24 |
         (val & 0x000000000000FF00) << 40 | (val & 0x00000000000000FF) << 56;
}

// Read 8 UC into a u64. Truncates UC if not char.
template <typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t
read8_to_u64(UC const *chars) {
  if (cpp20_and_in_constexpr() || !tinyobj_ff::is_same<UC, char>::value) {
    uint64_t val = 0;
    for (int i = 0; i < 8; ++i) {
      val |= uint64_t(uint8_t(*chars)) << (i * 8);
      ++chars;
    }
    return val;
  }
  uint64_t val;
  ::memcpy(&val, chars, sizeof(uint64_t));
#if FASTFLOAT_IS_BIG_ENDIAN == 1
  // Need to read as-if the number was in little-endian order.
  val = byteswap(val);
#endif
  return val;
}

#ifdef FASTFLOAT_SSE2

fastfloat_really_inline uint64_t simd_read8_to_u64(__m128i const data) {
  FASTFLOAT_SIMD_DISABLE_WARNINGS
  __m128i const packed = _mm_packus_epi16(data, data);
#ifdef FASTFLOAT_64BIT
  return uint64_t(_mm_cvtsi128_si64(packed));
#else
  uint64_t value;
  // Visual Studio + older versions of GCC don't support _mm_storeu_si64
  _mm_storel_epi64(reinterpret_cast<__m128i *>(&value), packed);
  return value;
#endif
  FASTFLOAT_SIMD_RESTORE_WARNINGS
}

fastfloat_really_inline uint64_t simd_read8_to_u64(char16_t const *chars) {
  FASTFLOAT_SIMD_DISABLE_WARNINGS
  return simd_read8_to_u64(
      _mm_loadu_si128(reinterpret_cast<__m128i const *>(chars)));
  FASTFLOAT_SIMD_RESTORE_WARNINGS
}

#elif defined(FASTFLOAT_NEON)

fastfloat_really_inline uint64_t simd_read8_to_u64(uint16x8_t const data) {
  FASTFLOAT_SIMD_DISABLE_WARNINGS
  uint8x8_t utf8_packed = vmovn_u16(data);
  return vget_lane_u64(vreinterpret_u64_u8(utf8_packed), 0);
  FASTFLOAT_SIMD_RESTORE_WARNINGS
}

fastfloat_really_inline uint64_t simd_read8_to_u64(char16_t const *chars) {
  FASTFLOAT_SIMD_DISABLE_WARNINGS
  return simd_read8_to_u64(
      vld1q_u16(reinterpret_cast<uint16_t const *>(chars)));
  FASTFLOAT_SIMD_RESTORE_WARNINGS
}

#endif // FASTFLOAT_SSE2

// MSVC SFINAE is broken pre-VS2017
#if defined(_MSC_VER) && _MSC_VER <= 1900
template <typename UC>
#else
template <typename UC, FASTFLOAT_ENABLE_IF(!has_simd_opt<UC>()) = 0>
#endif
// dummy for compile
uint64_t simd_read8_to_u64(UC const *) {
  return 0;
}

// credit  @aqrit
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 uint32_t
parse_eight_digits_unrolled(uint64_t val) {
  uint64_t const mask = 0x000000FF000000FF;
  uint64_t const mul1 = 0x000F424000000064; // 100 + (1000000ULL << 32)
  uint64_t const mul2 = 0x0000271000000001; // 1 + (10000ULL << 32)
  val -= 0x3030303030303030;
  val = (val * 10) + (val >> 8); // val = (val * 2561) >> 8;
  val = (((val & mask) * mul1) + (((val >> 16) & mask) * mul2)) >> 32;
  return uint32_t(val);
}

// Call this if chars are definitely 8 digits.
template <typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint32_t
parse_eight_digits_unrolled(UC const *chars) noexcept {
  if (cpp20_and_in_constexpr() || !has_simd_opt<UC>()) {
    return parse_eight_digits_unrolled(read8_to_u64(chars)); // truncation okay
  }
  return parse_eight_digits_unrolled(simd_read8_to_u64(chars));
}

// credit @aqrit
fastfloat_really_inline constexpr bool
is_made_of_eight_digits_fast(uint64_t val) noexcept {
  return !((((val + 0x4646464646464646) | (val - 0x3030303030303030)) &
            0x8080808080808080));
}

#ifdef FASTFLOAT_HAS_SIMD

// Call this if chars might not be 8 digits.
// Using this style (instead of is_made_of_eight_digits_fast() then
// parse_eight_digits_unrolled()) ensures we don't load SIMD registers twice.
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool
simd_parse_if_eight_digits_unrolled(char16_t const *chars,
                                    uint64_t &i) noexcept {
  if (cpp20_and_in_constexpr()) {
    return false;
  }
#ifdef FASTFLOAT_SSE2
  FASTFLOAT_SIMD_DISABLE_WARNINGS
  __m128i const data =
      _mm_loadu_si128(reinterpret_cast<__m128i const *>(chars));

  // (x - '0') <= 9
  // http://0x80.pl/articles/simd-parsing-int-sequences.html
  __m128i const t0 = _mm_add_epi16(data, _mm_set1_epi16(32720));
  __m128i const t1 = _mm_cmpgt_epi16(t0, _mm_set1_epi16(-32759));

  if (_mm_movemask_epi8(t1) == 0) {
    i = i * 100000000 + parse_eight_digits_unrolled(simd_read8_to_u64(data));
    return true;
  } else
    return false;
  FASTFLOAT_SIMD_RESTORE_WARNINGS
#elif defined(FASTFLOAT_NEON)
  FASTFLOAT_SIMD_DISABLE_WARNINGS
  uint16x8_t const data = vld1q_u16(reinterpret_cast<uint16_t const *>(chars));

  // (x - '0') <= 9
  // http://0x80.pl/articles/simd-parsing-int-sequences.html
  uint16x8_t const t0 = vsubq_u16(data, vmovq_n_u16('0'));
  uint16x8_t const mask = vcltq_u16(t0, vmovq_n_u16('9' - '0' + 1));

  if (vminvq_u16(mask) == 0xFFFF) {
    i = i * 100000000 + parse_eight_digits_unrolled(simd_read8_to_u64(data));
    return true;
  } else
    return false;
  FASTFLOAT_SIMD_RESTORE_WARNINGS
#else
  (void)chars;
  (void)i;
  return false;
#endif // FASTFLOAT_SSE2
}

#endif // FASTFLOAT_HAS_SIMD

// MSVC SFINAE is broken pre-VS2017
#if defined(_MSC_VER) && _MSC_VER <= 1900
template <typename UC>
#else
template <typename UC, FASTFLOAT_ENABLE_IF(!has_simd_opt<UC>()) = 0>
#endif
// dummy for compile
bool simd_parse_if_eight_digits_unrolled(UC const *, uint64_t &) {
  return 0;
}

template <typename UC, FASTFLOAT_ENABLE_IF(!tinyobj_ff::is_same<UC, char>::value) = 0>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
loop_parse_if_eight_digits(UC const *&p, UC const *const pend, uint64_t &i) {
  if (!has_simd_opt<UC>()) {
    return;
  }
  while ((tinyobj_ff::distance(p, pend) >= 8) &&
         simd_parse_if_eight_digits_unrolled(
             p, i)) { // in rare cases, this will overflow, but that's ok
    p += 8;
  }
}

fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
loop_parse_if_eight_digits(char const *&p, char const *const pend,
                           uint64_t &i) {
  // optimizes better than parse_if_eight_digits_unrolled() for UC = char.
  while ((tinyobj_ff::distance(p, pend) >= 8) &&
         is_made_of_eight_digits_fast(read8_to_u64(p))) {
    i = i * 100000000 +
        parse_eight_digits_unrolled(read8_to_u64(
            p)); // in rare cases, this will overflow, but that's ok
    p += 8;
  }
}

enum class parse_error {
  no_error,
  // [JSON-only] The minus sign must be followed by an integer.
  missing_integer_after_sign,
  // A sign must be followed by an integer or dot.
  missing_integer_or_dot_after_sign,
  // [JSON-only] The integer part must not have leading zeros.
  leading_zeros_in_integer_part,
  // [JSON-only] The integer part must have at least one digit.
  no_digits_in_integer_part,
  // [JSON-only] If there is a decimal point, there must be digits in the
  // fractional part.
  no_digits_in_fractional_part,
  // The mantissa must have at least one digit.
  no_digits_in_mantissa,
  // Scientific notation requires an exponential part.
  missing_exponential_part,
};

template <typename UC> struct parsed_number_string_t {
  int64_t exponent{0};
  uint64_t mantissa{0};
  UC const *lastmatch{nullptr};
  bool negative{false};
  bool valid{false};
  bool too_many_digits{false};
  // contains the range of the significant digits
  span<UC const> integer{};  // non-nullable
  span<UC const> fraction{}; // nullable
  parse_error error{parse_error::no_error};
};

using byte_span = span<char const>;
using parsed_number_string = parsed_number_string_t<char>;

template <typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 parsed_number_string_t<UC>
report_parse_error(UC const *p, parse_error error) {
  parsed_number_string_t<UC> answer;
  answer.valid = false;
  answer.lastmatch = p;
  answer.error = error;
  return answer;
}

// Assuming that you use no more than 19 digits, this will
// parse an ASCII string.
template <bool basic_json_fmt, typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 parsed_number_string_t<UC>
parse_number_string(UC const *p, UC const *pend,
                    parse_options_t<UC> options) noexcept {
  chars_format const fmt = detail::adjust_for_feature_macros(options.format);
  UC const decimal_point = options.decimal_point;

  parsed_number_string_t<UC> answer;
  answer.valid = false;
  answer.too_many_digits = false;
  // assume p < pend, so dereference without checks;
  answer.negative = (*p == UC('-'));
  // C++17 20.19.3.(7.1) explicitly forbids '+' sign here
  if ((*p == UC('-')) || (uint64_t(fmt & chars_format::allow_leading_plus) &&
                          !basic_json_fmt && *p == UC('+'))) {
    ++p;
    if (p == pend) {
      return report_parse_error<UC>(
          p, parse_error::missing_integer_or_dot_after_sign);
    }
    FASTFLOAT_IF_CONSTEXPR17(basic_json_fmt) {
      if (!is_integer(*p)) { // a sign must be followed by an integer
        return report_parse_error<UC>(p,
                                      parse_error::missing_integer_after_sign);
      }
    }
    else {
      if (!is_integer(*p) &&
          (*p !=
           decimal_point)) { // a sign must be followed by an integer or the dot
        return report_parse_error<UC>(
            p, parse_error::missing_integer_or_dot_after_sign);
      }
    }
  }
  UC const *const start_digits = p;

  uint64_t i = 0; // an unsigned int avoids signed overflows (which are bad)

  while ((p != pend) && is_integer(*p)) {
    // a multiplication by 10 is cheaper than an arbitrary integer
    // multiplication
    i = 10 * i +
        uint64_t(*p -
                 UC('0')); // might overflow, we will handle the overflow later
    ++p;
  }
  UC const *const end_of_integer_part = p;
  int64_t digit_count = int64_t(end_of_integer_part - start_digits);
  answer.integer = span<UC const>(start_digits, size_t(digit_count));
  FASTFLOAT_IF_CONSTEXPR17(basic_json_fmt) {
    // at least 1 digit in integer part, without leading zeros
    if (digit_count == 0) {
      return report_parse_error<UC>(p, parse_error::no_digits_in_integer_part);
    }
    if ((start_digits[0] == UC('0') && digit_count > 1)) {
      return report_parse_error<UC>(start_digits,
                                    parse_error::leading_zeros_in_integer_part);
    }
  }

  int64_t exponent = 0;
  bool const has_decimal_point = (p != pend) && (*p == decimal_point);
  if (has_decimal_point) {
    ++p;
    UC const *before = p;
    // can occur at most twice without overflowing, but let it occur more, since
    // for integers with many digits, digit parsing is the primary bottleneck.
    loop_parse_if_eight_digits(p, pend, i);

    while ((p != pend) && is_integer(*p)) {
      uint8_t digit = uint8_t(*p - UC('0'));
      ++p;
      i = i * 10 + digit; // in rare cases, this will overflow, but that's ok
    }
    exponent = before - p;
    answer.fraction = span<UC const>(before, size_t(p - before));
    digit_count -= exponent;
  }
  FASTFLOAT_IF_CONSTEXPR17(basic_json_fmt) {
    // at least 1 digit in fractional part
    if (has_decimal_point && exponent == 0) {
      return report_parse_error<UC>(p,
                                    parse_error::no_digits_in_fractional_part);
    }
  }
  else if (digit_count == 0) { // we must have encountered at least one integer!
    return report_parse_error<UC>(p, parse_error::no_digits_in_mantissa);
  }
  int64_t exp_number = 0; // explicit exponential part
  if ((uint64_t(fmt & chars_format::scientific) && (p != pend) &&
       ((UC('e') == *p) || (UC('E') == *p))) ||
      (uint64_t(fmt & detail::basic_fortran_fmt) && (p != pend) &&
       ((UC('+') == *p) || (UC('-') == *p) || (UC('d') == *p) ||
        (UC('D') == *p)))) {
    UC const *location_of_e = p;
    if ((UC('e') == *p) || (UC('E') == *p) || (UC('d') == *p) ||
        (UC('D') == *p)) {
      ++p;
    }
    bool neg_exp = false;
    if ((p != pend) && (UC('-') == *p)) {
      neg_exp = true;
      ++p;
    } else if ((p != pend) &&
               (UC('+') ==
                *p)) { // '+' on exponent is allowed by C++17 20.19.3.(7.1)
      ++p;
    }
    if ((p == pend) || !is_integer(*p)) {
      if (!uint64_t(fmt & chars_format::fixed)) {
        // The exponential part is invalid for scientific notation, so it must
        // be a trailing token for fixed notation. However, fixed notation is
        // disabled, so report a scientific notation error.
        return report_parse_error<UC>(p, parse_error::missing_exponential_part);
      }
      // Otherwise, we will be ignoring the 'e'.
      p = location_of_e;
    } else {
      while ((p != pend) && is_integer(*p)) {
        uint8_t digit = uint8_t(*p - UC('0'));
        if (exp_number < 0x10000000) {
          exp_number = 10 * exp_number + digit;
        }
        ++p;
      }
      if (neg_exp) {
        exp_number = -exp_number;
      }
      exponent += exp_number;
    }
  } else {
    // If it scientific and not fixed, we have to bail out.
    if (uint64_t(fmt & chars_format::scientific) &&
        !uint64_t(fmt & chars_format::fixed)) {
      return report_parse_error<UC>(p, parse_error::missing_exponential_part);
    }
  }
  answer.lastmatch = p;
  answer.valid = true;

  // If we frequently had to deal with long strings of digits,
  // we could extend our code by using a 128-bit integer instead
  // of a 64-bit integer. However, this is uncommon.
  //
  // We can deal with up to 19 digits.
  if (digit_count > 19) { // this is uncommon
    // It is possible that the integer had an overflow.
    // We have to handle the case where we have 0.0000somenumber.
    // We need to be mindful of the case where we only have zeroes...
    // E.g., 0.000000000...000.
    UC const *start = start_digits;
    while ((start != pend) && (*start == UC('0') || *start == decimal_point)) {
      if (*start == UC('0')) {
        digit_count--;
      }
      start++;
    }

    if (digit_count > 19) {
      answer.too_many_digits = true;
      // Let us start again, this time, avoiding overflows.
      // We don't need to check if is_integer, since we use the
      // pre-tokenized spans from above.
      i = 0;
      p = answer.integer.ptr;
      UC const *int_end = p + answer.integer.len();
      uint64_t const minimal_nineteen_digit_integer{1000000000000000000};
      while ((i < minimal_nineteen_digit_integer) && (p != int_end)) {
        i = i * 10 + uint64_t(*p - UC('0'));
        ++p;
      }
      if (i >= minimal_nineteen_digit_integer) { // We have a big integers
        exponent = end_of_integer_part - p + exp_number;
      } else { // We have a value with a fractional component.
        p = answer.fraction.ptr;
        UC const *frac_end = p + answer.fraction.len();
        while ((i < minimal_nineteen_digit_integer) && (p != frac_end)) {
          i = i * 10 + uint64_t(*p - UC('0'));
          ++p;
        }
        exponent = answer.fraction.ptr - p + exp_number;
      }
      // We have now corrected both exponent and i, to a truncated value
    }
  }
  answer.exponent = exponent;
  answer.mantissa = i;
  return answer;
}

template <typename T, typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
parse_int_string(UC const *p, UC const *pend, T &value,
                 parse_options_t<UC> options) {
  chars_format const fmt = detail::adjust_for_feature_macros(options.format);
  int const base = options.base;

  from_chars_result_t<UC> answer;

  UC const *const first = p;

  bool const negative = (*p == UC('-'));
#ifdef FASTFLOAT_VISUAL_STUDIO
#pragma warning(push)
#pragma warning(disable : 4127)
#endif
  if (!tinyobj_ff::is_signed<T>::value && negative) {
#ifdef FASTFLOAT_VISUAL_STUDIO
#pragma warning(pop)
#endif
    answer.ec = tinyobj_ff::ff_errc::invalid_argument;
    answer.ptr = first;
    return answer;
  }
  if ((*p == UC('-')) ||
      (uint64_t(fmt & chars_format::allow_leading_plus) && (*p == UC('+')))) {
    ++p;
  }

  UC const *const start_num = p;

  while (p != pend && *p == UC('0')) {
    ++p;
  }

  bool const has_leading_zeros = p > start_num;

  UC const *const start_digits = p;

  uint64_t i = 0;
  if (base == 10) {
    loop_parse_if_eight_digits(p, pend, i); // use SIMD if possible
  }
  while (p != pend) {
    uint8_t digit = ch_to_digit(*p);
    if (digit >= base) {
      break;
    }
    i = uint64_t(base) * i + digit; // might overflow, check this later
    p++;
  }

  size_t digit_count = size_t(p - start_digits);

  if (digit_count == 0) {
    if (has_leading_zeros) {
      value = 0;
      answer.ec = tinyobj_ff::ff_errc();
      answer.ptr = p;
    } else {
      answer.ec = tinyobj_ff::ff_errc::invalid_argument;
      answer.ptr = first;
    }
    return answer;
  }

  answer.ptr = p;

  // check u64 overflow
  size_t max_digits = max_digits_u64(base);
  if (digit_count > max_digits) {
    answer.ec = tinyobj_ff::ff_errc::result_out_of_range;
    return answer;
  }
  // this check can be eliminated for all other types, but they will all require
  // a max_digits(base) equivalent
  if (digit_count == max_digits && i < min_safe_u64(base)) {
    answer.ec = tinyobj_ff::ff_errc::result_out_of_range;
    return answer;
  }

  // check other types overflow
  if (!tinyobj_ff::is_same<T, uint64_t>::value) {
    if (i > uint64_t(std::numeric_limits<T>::max()) + uint64_t(negative)) {
      answer.ec = tinyobj_ff::ff_errc::result_out_of_range;
      return answer;
    }
  }

  if (negative) {
#ifdef FASTFLOAT_VISUAL_STUDIO
#pragma warning(push)
#pragma warning(disable : 4146)
#endif
    // this weird workaround is required because:
    // - converting unsigned to signed when its value is greater than signed max
    // is UB pre-C++23.
    // - reinterpret_casting (~i + 1) would work, but it is not constexpr
    // this is always optimized into a neg instruction (note: T is an integer
    // type)
    value = T(-std::numeric_limits<T>::max() -
              T(i - uint64_t(std::numeric_limits<T>::max())));
#ifdef FASTFLOAT_VISUAL_STUDIO
#pragma warning(pop)
#endif
  } else {
    value = T(i);
  }

  answer.ec = tinyobj_ff::ff_errc();
  return answer;
}

} // namespace fast_float

#endif

#ifndef FASTFLOAT_FAST_TABLE_H
#define FASTFLOAT_FAST_TABLE_H

namespace fast_float {

/**
 * When mapping numbers from decimal to binary,
 * we go from w * 10^q to m * 2^p but we have
 * 10^q = 5^q * 2^q, so effectively
 * we are trying to match
 * w * 2^q * 5^q to m * 2^p. Thus the powers of two
 * are not a concern since they can be represented
 * exactly using the binary notation, only the powers of five
 * affect the binary significand.
 */

/**
 * The smallest non-zero float (binary64) is 2^-1074.
 * We take as input numbers of the form w x 10^q where w < 2^64.
 * We have that w * 10^-343  <  2^(64-344) 5^-343 < 2^-1076.
 * However, we have that
 * (2^64-1) * 10^-342 =  (2^64-1) * 2^-342 * 5^-342 > 2^-1074.
 * Thus it is possible for a number of the form w * 10^-342 where
 * w is a 64-bit value to be a non-zero floating-point number.
 *********
 * Any number of form w * 10^309 where w>= 1 is going to be
 * infinite in binary64 so we never need to worry about powers
 * of 5 greater than 308.
 */
template <class unused = void> struct powers_template {

  constexpr static int smallest_power_of_five =
      binary_format<double>::smallest_power_of_ten();
  constexpr static int largest_power_of_five =
      binary_format<double>::largest_power_of_ten();
  constexpr static int number_of_entries =
      2 * (largest_power_of_five - smallest_power_of_five + 1);
  // Powers of five from 5^-342 all the way to 5^308 rounded toward one.
  constexpr static uint64_t power_of_five_128[number_of_entries] = {
      0xeef453d6923bd65a, 0x113faa2906a13b3f,
      0x9558b4661b6565f8, 0x4ac7ca59a424c507,
      0xbaaee17fa23ebf76, 0x5d79bcf00d2df649,
      0xe95a99df8ace6f53, 0xf4d82c2c107973dc,
      0x91d8a02bb6c10594, 0x79071b9b8a4be869,
      0xb64ec836a47146f9, 0x9748e2826cdee284,
      0xe3e27a444d8d98b7, 0xfd1b1b2308169b25,
      0x8e6d8c6ab0787f72, 0xfe30f0f5e50e20f7,
      0xb208ef855c969f4f, 0xbdbd2d335e51a935,
      0xde8b2b66b3bc4723, 0xad2c788035e61382,
      0x8b16fb203055ac76, 0x4c3bcb5021afcc31,
      0xaddcb9e83c6b1793, 0xdf4abe242a1bbf3d,
      0xd953e8624b85dd78, 0xd71d6dad34a2af0d,
      0x87d4713d6f33aa6b, 0x8672648c40e5ad68,
      0xa9c98d8ccb009506, 0x680efdaf511f18c2,
      0xd43bf0effdc0ba48, 0x212bd1b2566def2,
      0x84a57695fe98746d, 0x14bb630f7604b57,
      0xa5ced43b7e3e9188, 0x419ea3bd35385e2d,
      0xcf42894a5dce35ea, 0x52064cac828675b9,
      0x818995ce7aa0e1b2, 0x7343efebd1940993,
      0xa1ebfb4219491a1f, 0x1014ebe6c5f90bf8,
      0xca66fa129f9b60a6, 0xd41a26e077774ef6,
      0xfd00b897478238d0, 0x8920b098955522b4,
      0x9e20735e8cb16382, 0x55b46e5f5d5535b0,
      0xc5a890362fddbc62, 0xeb2189f734aa831d,
      0xf712b443bbd52b7b, 0xa5e9ec7501d523e4,
      0x9a6bb0aa55653b2d, 0x47b233c92125366e,
      0xc1069cd4eabe89f8, 0x999ec0bb696e840a,
      0xf148440a256e2c76, 0xc00670ea43ca250d,
      0x96cd2a865764dbca, 0x380406926a5e5728,
      0xbc807527ed3e12bc, 0xc605083704f5ecf2,
      0xeba09271e88d976b, 0xf7864a44c633682e,
      0x93445b8731587ea3, 0x7ab3ee6afbe0211d,
      0xb8157268fdae9e4c, 0x5960ea05bad82964,
      0xe61acf033d1a45df, 0x6fb92487298e33bd,
      0x8fd0c16206306bab, 0xa5d3b6d479f8e056,
      0xb3c4f1ba87bc8696, 0x8f48a4899877186c,
      0xe0b62e2929aba83c, 0x331acdabfe94de87,
      0x8c71dcd9ba0b4925, 0x9ff0c08b7f1d0b14,
      0xaf8e5410288e1b6f, 0x7ecf0ae5ee44dd9,
      0xdb71e91432b1a24a, 0xc9e82cd9f69d6150,
      0x892731ac9faf056e, 0xbe311c083a225cd2,
      0xab70fe17c79ac6ca, 0x6dbd630a48aaf406,
      0xd64d3d9db981787d, 0x92cbbccdad5b108,
      0x85f0468293f0eb4e, 0x25bbf56008c58ea5,
      0xa76c582338ed2621, 0xaf2af2b80af6f24e,
      0xd1476e2c07286faa, 0x1af5af660db4aee1,
      0x82cca4db847945ca, 0x50d98d9fc890ed4d,
      0xa37fce126597973c, 0xe50ff107bab528a0,
      0xcc5fc196fefd7d0c, 0x1e53ed49a96272c8,
      0xff77b1fcbebcdc4f, 0x25e8e89c13bb0f7a,
      0x9faacf3df73609b1, 0x77b191618c54e9ac,
      0xc795830d75038c1d, 0xd59df5b9ef6a2417,
      0xf97ae3d0d2446f25, 0x4b0573286b44ad1d,
      0x9becce62836ac577, 0x4ee367f9430aec32,
      0xc2e801fb244576d5, 0x229c41f793cda73f,
      0xf3a20279ed56d48a, 0x6b43527578c1110f,
      0x9845418c345644d6, 0x830a13896b78aaa9,
      0xbe5691ef416bd60c, 0x23cc986bc656d553,
      0xedec366b11c6cb8f, 0x2cbfbe86b7ec8aa8,
      0x94b3a202eb1c3f39, 0x7bf7d71432f3d6a9,
      0xb9e08a83a5e34f07, 0xdaf5ccd93fb0cc53,
      0xe858ad248f5c22c9, 0xd1b3400f8f9cff68,
      0x91376c36d99995be, 0x23100809b9c21fa1,
      0xb58547448ffffb2d, 0xabd40a0c2832a78a,
      0xe2e69915b3fff9f9, 0x16c90c8f323f516c,
      0x8dd01fad907ffc3b, 0xae3da7d97f6792e3,
      0xb1442798f49ffb4a, 0x99cd11cfdf41779c,
      0xdd95317f31c7fa1d, 0x40405643d711d583,
      0x8a7d3eef7f1cfc52, 0x482835ea666b2572,
      0xad1c8eab5ee43b66, 0xda3243650005eecf,
      0xd863b256369d4a40, 0x90bed43e40076a82,
      0x873e4f75e2224e68, 0x5a7744a6e804a291,
      0xa90de3535aaae202, 0x711515d0a205cb36,
      0xd3515c2831559a83, 0xd5a5b44ca873e03,
      0x8412d9991ed58091, 0xe858790afe9486c2,
      0xa5178fff668ae0b6, 0x626e974dbe39a872,
      0xce5d73ff402d98e3, 0xfb0a3d212dc8128f,
      0x80fa687f881c7f8e, 0x7ce66634bc9d0b99,
      0xa139029f6a239f72, 0x1c1fffc1ebc44e80,
      0xc987434744ac874e, 0xa327ffb266b56220,
      0xfbe9141915d7a922, 0x4bf1ff9f0062baa8,
      0x9d71ac8fada6c9b5, 0x6f773fc3603db4a9,
      0xc4ce17b399107c22, 0xcb550fb4384d21d3,
      0xf6019da07f549b2b, 0x7e2a53a146606a48,
      0x99c102844f94e0fb, 0x2eda7444cbfc426d,
      0xc0314325637a1939, 0xfa911155fefb5308,
      0xf03d93eebc589f88, 0x793555ab7eba27ca,
      0x96267c7535b763b5, 0x4bc1558b2f3458de,
      0xbbb01b9283253ca2, 0x9eb1aaedfb016f16,
      0xea9c227723ee8bcb, 0x465e15a979c1cadc,
      0x92a1958a7675175f, 0xbfacd89ec191ec9,
      0xb749faed14125d36, 0xcef980ec671f667b,
      0xe51c79a85916f484, 0x82b7e12780e7401a,
      0x8f31cc0937ae58d2, 0xd1b2ecb8b0908810,
      0xb2fe3f0b8599ef07, 0x861fa7e6dcb4aa15,
      0xdfbdcece67006ac9, 0x67a791e093e1d49a,
      0x8bd6a141006042bd, 0xe0c8bb2c5c6d24e0,
      0xaecc49914078536d, 0x58fae9f773886e18,
      0xda7f5bf590966848, 0xaf39a475506a899e,
      0x888f99797a5e012d, 0x6d8406c952429603,
      0xaab37fd7d8f58178, 0xc8e5087ba6d33b83,
      0xd5605fcdcf32e1d6, 0xfb1e4a9a90880a64,
      0x855c3be0a17fcd26, 0x5cf2eea09a55067f,
      0xa6b34ad8c9dfc06f, 0xf42faa48c0ea481e,
      0xd0601d8efc57b08b, 0xf13b94daf124da26,
      0x823c12795db6ce57, 0x76c53d08d6b70858,
      0xa2cb1717b52481ed, 0x54768c4b0c64ca6e,
      0xcb7ddcdda26da268, 0xa9942f5dcf7dfd09,
      0xfe5d54150b090b02, 0xd3f93b35435d7c4c,
      0x9efa548d26e5a6e1, 0xc47bc5014a1a6daf,
      0xc6b8e9b0709f109a, 0x359ab6419ca1091b,
      0xf867241c8cc6d4c0, 0xc30163d203c94b62,
      0x9b407691d7fc44f8, 0x79e0de63425dcf1d,
      0xc21094364dfb5636, 0x985915fc12f542e4,
      0xf294b943e17a2bc4, 0x3e6f5b7b17b2939d,
      0x979cf3ca6cec5b5a, 0xa705992ceecf9c42,
      0xbd8430bd08277231, 0x50c6ff782a838353,
      0xece53cec4a314ebd, 0xa4f8bf5635246428,
      0x940f4613ae5ed136, 0x871b7795e136be99,
      0xb913179899f68584, 0x28e2557b59846e3f,
      0xe757dd7ec07426e5, 0x331aeada2fe589cf,
      0x9096ea6f3848984f, 0x3ff0d2c85def7621,
      0xb4bca50b065abe63, 0xfed077a756b53a9,
      0xe1ebce4dc7f16dfb, 0xd3e8495912c62894,
      0x8d3360f09cf6e4bd, 0x64712dd7abbbd95c,
      0xb080392cc4349dec, 0xbd8d794d96aacfb3,
      0xdca04777f541c567, 0xecf0d7a0fc5583a0,
      0x89e42caaf9491b60, 0xf41686c49db57244,
      0xac5d37d5b79b6239, 0x311c2875c522ced5,
      0xd77485cb25823ac7, 0x7d633293366b828b,
      0x86a8d39ef77164bc, 0xae5dff9c02033197,
      0xa8530886b54dbdeb, 0xd9f57f830283fdfc,
      0xd267caa862a12d66, 0xd072df63c324fd7b,
      0x8380dea93da4bc60, 0x4247cb9e59f71e6d,
      0xa46116538d0deb78, 0x52d9be85f074e608,
      0xcd795be870516656, 0x67902e276c921f8b,
      0x806bd9714632dff6, 0xba1cd8a3db53b6,
      0xa086cfcd97bf97f3, 0x80e8a40eccd228a4,
      0xc8a883c0fdaf7df0, 0x6122cd128006b2cd,
      0xfad2a4b13d1b5d6c, 0x796b805720085f81,
      0x9cc3a6eec6311a63, 0xcbe3303674053bb0,
      0xc3f490aa77bd60fc, 0xbedbfc4411068a9c,
      0xf4f1b4d515acb93b, 0xee92fb5515482d44,
      0x991711052d8bf3c5, 0x751bdd152d4d1c4a,
      0xbf5cd54678eef0b6, 0xd262d45a78a0635d,
      0xef340a98172aace4, 0x86fb897116c87c34,
      0x9580869f0e7aac0e, 0xd45d35e6ae3d4da0,
      0xbae0a846d2195712, 0x8974836059cca109,
      0xe998d258869facd7, 0x2bd1a438703fc94b,
      0x91ff83775423cc06, 0x7b6306a34627ddcf,
      0xb67f6455292cbf08, 0x1a3bc84c17b1d542,
      0xe41f3d6a7377eeca, 0x20caba5f1d9e4a93,
      0x8e938662882af53e, 0x547eb47b7282ee9c,
      0xb23867fb2a35b28d, 0xe99e619a4f23aa43,
      0xdec681f9f4c31f31, 0x6405fa00e2ec94d4,
      0x8b3c113c38f9f37e, 0xde83bc408dd3dd04,
      0xae0b158b4738705e, 0x9624ab50b148d445,
      0xd98ddaee19068c76, 0x3badd624dd9b0957,
      0x87f8a8d4cfa417c9, 0xe54ca5d70a80e5d6,
      0xa9f6d30a038d1dbc, 0x5e9fcf4ccd211f4c,
      0xd47487cc8470652b, 0x7647c3200069671f,
      0x84c8d4dfd2c63f3b, 0x29ecd9f40041e073,
      0xa5fb0a17c777cf09, 0xf468107100525890,
      0xcf79cc9db955c2cc, 0x7182148d4066eeb4,
      0x81ac1fe293d599bf, 0xc6f14cd848405530,
      0xa21727db38cb002f, 0xb8ada00e5a506a7c,
      0xca9cf1d206fdc03b, 0xa6d90811f0e4851c,
      0xfd442e4688bd304a, 0x908f4a166d1da663,
      0x9e4a9cec15763e2e, 0x9a598e4e043287fe,
      0xc5dd44271ad3cdba, 0x40eff1e1853f29fd,
      0xf7549530e188c128, 0xd12bee59e68ef47c,
      0x9a94dd3e8cf578b9, 0x82bb74f8301958ce,
      0xc13a148e3032d6e7, 0xe36a52363c1faf01,
      0xf18899b1bc3f8ca1, 0xdc44e6c3cb279ac1,
      0x96f5600f15a7b7e5, 0x29ab103a5ef8c0b9,
      0xbcb2b812db11a5de, 0x7415d448f6b6f0e7,
      0xebdf661791d60f56, 0x111b495b3464ad21,
      0x936b9fcebb25c995, 0xcab10dd900beec34,
      0xb84687c269ef3bfb, 0x3d5d514f40eea742,
      0xe65829b3046b0afa, 0xcb4a5a3112a5112,
      0x8ff71a0fe2c2e6dc, 0x47f0e785eaba72ab,
      0xb3f4e093db73a093, 0x59ed216765690f56,
      0xe0f218b8d25088b8, 0x306869c13ec3532c,
      0x8c974f7383725573, 0x1e414218c73a13fb,
      0xafbd2350644eeacf, 0xe5d1929ef90898fa,
      0xdbac6c247d62a583, 0xdf45f746b74abf39,
      0x894bc396ce5da772, 0x6b8bba8c328eb783,
      0xab9eb47c81f5114f, 0x66ea92f3f326564,
      0xd686619ba27255a2, 0xc80a537b0efefebd,
      0x8613fd0145877585, 0xbd06742ce95f5f36,
      0xa798fc4196e952e7, 0x2c48113823b73704,
      0xd17f3b51fca3a7a0, 0xf75a15862ca504c5,
      0x82ef85133de648c4, 0x9a984d73dbe722fb,
      0xa3ab66580d5fdaf5, 0xc13e60d0d2e0ebba,
      0xcc963fee10b7d1b3, 0x318df905079926a8,
      0xffbbcfe994e5c61f, 0xfdf17746497f7052,
      0x9fd561f1fd0f9bd3, 0xfeb6ea8bedefa633,
      0xc7caba6e7c5382c8, 0xfe64a52ee96b8fc0,
      0xf9bd690a1b68637b, 0x3dfdce7aa3c673b0,
      0x9c1661a651213e2d, 0x6bea10ca65c084e,
      0xc31bfa0fe5698db8, 0x486e494fcff30a62,
      0xf3e2f893dec3f126, 0x5a89dba3c3efccfa,
      0x986ddb5c6b3a76b7, 0xf89629465a75e01c,
      0xbe89523386091465, 0xf6bbb397f1135823,
      0xee2ba6c0678b597f, 0x746aa07ded582e2c,
      0x94db483840b717ef, 0xa8c2a44eb4571cdc,
      0xba121a4650e4ddeb, 0x92f34d62616ce413,
      0xe896a0d7e51e1566, 0x77b020baf9c81d17,
      0x915e2486ef32cd60, 0xace1474dc1d122e,
      0xb5b5ada8aaff80b8, 0xd819992132456ba,
      0xe3231912d5bf60e6, 0x10e1fff697ed6c69,
      0x8df5efabc5979c8f, 0xca8d3ffa1ef463c1,
      0xb1736b96b6fd83b3, 0xbd308ff8a6b17cb2,
      0xddd0467c64bce4a0, 0xac7cb3f6d05ddbde,
      0x8aa22c0dbef60ee4, 0x6bcdf07a423aa96b,
      0xad4ab7112eb3929d, 0x86c16c98d2c953c6,
      0xd89d64d57a607744, 0xe871c7bf077ba8b7,
      0x87625f056c7c4a8b, 0x11471cd764ad4972,
      0xa93af6c6c79b5d2d, 0xd598e40d3dd89bcf,
      0xd389b47879823479, 0x4aff1d108d4ec2c3,
      0x843610cb4bf160cb, 0xcedf722a585139ba,
      0xa54394fe1eedb8fe, 0xc2974eb4ee658828,
      0xce947a3da6a9273e, 0x733d226229feea32,
      0x811ccc668829b887, 0x806357d5a3f525f,
      0xa163ff802a3426a8, 0xca07c2dcb0cf26f7,
      0xc9bcff6034c13052, 0xfc89b393dd02f0b5,
      0xfc2c3f3841f17c67, 0xbbac2078d443ace2,
      0x9d9ba7832936edc0, 0xd54b944b84aa4c0d,
      0xc5029163f384a931, 0xa9e795e65d4df11,
      0xf64335bcf065d37d, 0x4d4617b5ff4a16d5,
      0x99ea0196163fa42e, 0x504bced1bf8e4e45,
      0xc06481fb9bcf8d39, 0xe45ec2862f71e1d6,
      0xf07da27a82c37088, 0x5d767327bb4e5a4c,
      0x964e858c91ba2655, 0x3a6a07f8d510f86f,
      0xbbe226efb628afea, 0x890489f70a55368b,
      0xeadab0aba3b2dbe5, 0x2b45ac74ccea842e,
      0x92c8ae6b464fc96f, 0x3b0b8bc90012929d,
      0xb77ada0617e3bbcb, 0x9ce6ebb40173744,
      0xe55990879ddcaabd, 0xcc420a6a101d0515,
      0x8f57fa54c2a9eab6, 0x9fa946824a12232d,
      0xb32df8e9f3546564, 0x47939822dc96abf9,
      0xdff9772470297ebd, 0x59787e2b93bc56f7,
      0x8bfbea76c619ef36, 0x57eb4edb3c55b65a,
      0xaefae51477a06b03, 0xede622920b6b23f1,
      0xdab99e59958885c4, 0xe95fab368e45eced,
      0x88b402f7fd75539b, 0x11dbcb0218ebb414,
      0xaae103b5fcd2a881, 0xd652bdc29f26a119,
      0xd59944a37c0752a2, 0x4be76d3346f0495f,
      0x857fcae62d8493a5, 0x6f70a4400c562ddb,
      0xa6dfbd9fb8e5b88e, 0xcb4ccd500f6bb952,
      0xd097ad07a71f26b2, 0x7e2000a41346a7a7,
      0x825ecc24c873782f, 0x8ed400668c0c28c8,
      0xa2f67f2dfa90563b, 0x728900802f0f32fa,
      0xcbb41ef979346bca, 0x4f2b40a03ad2ffb9,
      0xfea126b7d78186bc, 0xe2f610c84987bfa8,
      0x9f24b832e6b0f436, 0xdd9ca7d2df4d7c9,
      0xc6ede63fa05d3143, 0x91503d1c79720dbb,
      0xf8a95fcf88747d94, 0x75a44c6397ce912a,
      0x9b69dbe1b548ce7c, 0xc986afbe3ee11aba,
      0xc24452da229b021b, 0xfbe85badce996168,
      0xf2d56790ab41c2a2, 0xfae27299423fb9c3,
      0x97c560ba6b0919a5, 0xdccd879fc967d41a,
      0xbdb6b8e905cb600f, 0x5400e987bbc1c920,
      0xed246723473e3813, 0x290123e9aab23b68,
      0x9436c0760c86e30b, 0xf9a0b6720aaf6521,
      0xb94470938fa89bce, 0xf808e40e8d5b3e69,
      0xe7958cb87392c2c2, 0xb60b1d1230b20e04,
      0x90bd77f3483bb9b9, 0xb1c6f22b5e6f48c2,
      0xb4ecd5f01a4aa828, 0x1e38aeb6360b1af3,
      0xe2280b6c20dd5232, 0x25c6da63c38de1b0,
      0x8d590723948a535f, 0x579c487e5a38ad0e,
      0xb0af48ec79ace837, 0x2d835a9df0c6d851,
      0xdcdb1b2798182244, 0xf8e431456cf88e65,
      0x8a08f0f8bf0f156b, 0x1b8e9ecb641b58ff,
      0xac8b2d36eed2dac5, 0xe272467e3d222f3f,
      0xd7adf884aa879177, 0x5b0ed81dcc6abb0f,
      0x86ccbb52ea94baea, 0x98e947129fc2b4e9,
      0xa87fea27a539e9a5, 0x3f2398d747b36224,
      0xd29fe4b18e88640e, 0x8eec7f0d19a03aad,
      0x83a3eeeef9153e89, 0x1953cf68300424ac,
      0xa48ceaaab75a8e2b, 0x5fa8c3423c052dd7,
      0xcdb02555653131b6, 0x3792f412cb06794d,
      0x808e17555f3ebf11, 0xe2bbd88bbee40bd0,
      0xa0b19d2ab70e6ed6, 0x5b6aceaeae9d0ec4,
      0xc8de047564d20a8b, 0xf245825a5a445275,
      0xfb158592be068d2e, 0xeed6e2f0f0d56712,
      0x9ced737bb6c4183d, 0x55464dd69685606b,
      0xc428d05aa4751e4c, 0xaa97e14c3c26b886,
      0xf53304714d9265df, 0xd53dd99f4b3066a8,
      0x993fe2c6d07b7fab, 0xe546a8038efe4029,
      0xbf8fdb78849a5f96, 0xde98520472bdd033,
      0xef73d256a5c0f77c, 0x963e66858f6d4440,
      0x95a8637627989aad, 0xdde7001379a44aa8,
      0xbb127c53b17ec159, 0x5560c018580d5d52,
      0xe9d71b689dde71af, 0xaab8f01e6e10b4a6,
      0x9226712162ab070d, 0xcab3961304ca70e8,
      0xb6b00d69bb55c8d1, 0x3d607b97c5fd0d22,
      0xe45c10c42a2b3b05, 0x8cb89a7db77c506a,
      0x8eb98a7a9a5b04e3, 0x77f3608e92adb242,
      0xb267ed1940f1c61c, 0x55f038b237591ed3,
      0xdf01e85f912e37a3, 0x6b6c46dec52f6688,
      0x8b61313bbabce2c6, 0x2323ac4b3b3da015,
      0xae397d8aa96c1b77, 0xabec975e0a0d081a,
      0xd9c7dced53c72255, 0x96e7bd358c904a21,
      0x881cea14545c7575, 0x7e50d64177da2e54,
      0xaa242499697392d2, 0xdde50bd1d5d0b9e9,
      0xd4ad2dbfc3d07787, 0x955e4ec64b44e864,
      0x84ec3c97da624ab4, 0xbd5af13bef0b113e,
      0xa6274bbdd0fadd61, 0xecb1ad8aeacdd58e,
      0xcfb11ead453994ba, 0x67de18eda5814af2,
      0x81ceb32c4b43fcf4, 0x80eacf948770ced7,
      0xa2425ff75e14fc31, 0xa1258379a94d028d,
      0xcad2f7f5359a3b3e, 0x96ee45813a04330,
      0xfd87b5f28300ca0d, 0x8bca9d6e188853fc,
      0x9e74d1b791e07e48, 0x775ea264cf55347e,
      0xc612062576589dda, 0x95364afe032a819e,
      0xf79687aed3eec551, 0x3a83ddbd83f52205,
      0x9abe14cd44753b52, 0xc4926a9672793543,
      0xc16d9a0095928a27, 0x75b7053c0f178294,
      0xf1c90080baf72cb1, 0x5324c68b12dd6339,
      0x971da05074da7bee, 0xd3f6fc16ebca5e04,
      0xbce5086492111aea, 0x88f4bb1ca6bcf585,
      0xec1e4a7db69561a5, 0x2b31e9e3d06c32e6,
      0x9392ee8e921d5d07, 0x3aff322e62439fd0,
      0xb877aa3236a4b449, 0x9befeb9fad487c3,
      0xe69594bec44de15b, 0x4c2ebe687989a9b4,
      0x901d7cf73ab0acd9, 0xf9d37014bf60a11,
      0xb424dc35095cd80f, 0x538484c19ef38c95,
      0xe12e13424bb40e13, 0x2865a5f206b06fba,
      0x8cbccc096f5088cb, 0xf93f87b7442e45d4,
      0xafebff0bcb24aafe, 0xf78f69a51539d749,
      0xdbe6fecebdedd5be, 0xb573440e5a884d1c,
      0x89705f4136b4a597, 0x31680a88f8953031,
      0xabcc77118461cefc, 0xfdc20d2b36ba7c3e,
      0xd6bf94d5e57a42bc, 0x3d32907604691b4d,
      0x8637bd05af6c69b5, 0xa63f9a49c2c1b110,
      0xa7c5ac471b478423, 0xfcf80dc33721d54,
      0xd1b71758e219652b, 0xd3c36113404ea4a9,
      0x83126e978d4fdf3b, 0x645a1cac083126ea,
      0xa3d70a3d70a3d70a, 0x3d70a3d70a3d70a4,
      0xcccccccccccccccc, 0xcccccccccccccccd,
      0x8000000000000000, 0x0,
      0xa000000000000000, 0x0,
      0xc800000000000000, 0x0,
      0xfa00000000000000, 0x0,
      0x9c40000000000000, 0x0,
      0xc350000000000000, 0x0,
      0xf424000000000000, 0x0,
      0x9896800000000000, 0x0,
      0xbebc200000000000, 0x0,
      0xee6b280000000000, 0x0,
      0x9502f90000000000, 0x0,
      0xba43b74000000000, 0x0,
      0xe8d4a51000000000, 0x0,
      0x9184e72a00000000, 0x0,
      0xb5e620f480000000, 0x0,
      0xe35fa931a0000000, 0x0,
      0x8e1bc9bf04000000, 0x0,
      0xb1a2bc2ec5000000, 0x0,
      0xde0b6b3a76400000, 0x0,
      0x8ac7230489e80000, 0x0,
      0xad78ebc5ac620000, 0x0,
      0xd8d726b7177a8000, 0x0,
      0x878678326eac9000, 0x0,
      0xa968163f0a57b400, 0x0,
      0xd3c21bcecceda100, 0x0,
      0x84595161401484a0, 0x0,
      0xa56fa5b99019a5c8, 0x0,
      0xcecb8f27f4200f3a, 0x0,
      0x813f3978f8940984, 0x4000000000000000,
      0xa18f07d736b90be5, 0x5000000000000000,
      0xc9f2c9cd04674ede, 0xa400000000000000,
      0xfc6f7c4045812296, 0x4d00000000000000,
      0x9dc5ada82b70b59d, 0xf020000000000000,
      0xc5371912364ce305, 0x6c28000000000000,
      0xf684df56c3e01bc6, 0xc732000000000000,
      0x9a130b963a6c115c, 0x3c7f400000000000,
      0xc097ce7bc90715b3, 0x4b9f100000000000,
      0xf0bdc21abb48db20, 0x1e86d40000000000,
      0x96769950b50d88f4, 0x1314448000000000,
      0xbc143fa4e250eb31, 0x17d955a000000000,
      0xeb194f8e1ae525fd, 0x5dcfab0800000000,
      0x92efd1b8d0cf37be, 0x5aa1cae500000000,
      0xb7abc627050305ad, 0xf14a3d9e40000000,
      0xe596b7b0c643c719, 0x6d9ccd05d0000000,
      0x8f7e32ce7bea5c6f, 0xe4820023a2000000,
      0xb35dbf821ae4f38b, 0xdda2802c8a800000,
      0xe0352f62a19e306e, 0xd50b2037ad200000,
      0x8c213d9da502de45, 0x4526f422cc340000,
      0xaf298d050e4395d6, 0x9670b12b7f410000,
      0xdaf3f04651d47b4c, 0x3c0cdd765f114000,
      0x88d8762bf324cd0f, 0xa5880a69fb6ac800,
      0xab0e93b6efee0053, 0x8eea0d047a457a00,
      0xd5d238a4abe98068, 0x72a4904598d6d880,
      0x85a36366eb71f041, 0x47a6da2b7f864750,
      0xa70c3c40a64e6c51, 0x999090b65f67d924,
      0xd0cf4b50cfe20765, 0xfff4b4e3f741cf6d,
      0x82818f1281ed449f, 0xbff8f10e7a8921a4,
      0xa321f2d7226895c7, 0xaff72d52192b6a0d,
      0xcbea6f8ceb02bb39, 0x9bf4f8a69f764490,
      0xfee50b7025c36a08, 0x2f236d04753d5b4,
      0x9f4f2726179a2245, 0x1d762422c946590,
      0xc722f0ef9d80aad6, 0x424d3ad2b7b97ef5,
      0xf8ebad2b84e0d58b, 0xd2e0898765a7deb2,
      0x9b934c3b330c8577, 0x63cc55f49f88eb2f,
      0xc2781f49ffcfa6d5, 0x3cbf6b71c76b25fb,
      0xf316271c7fc3908a, 0x8bef464e3945ef7a,
      0x97edd871cfda3a56, 0x97758bf0e3cbb5ac,
      0xbde94e8e43d0c8ec, 0x3d52eeed1cbea317,
      0xed63a231d4c4fb27, 0x4ca7aaa863ee4bdd,
      0x945e455f24fb1cf8, 0x8fe8caa93e74ef6a,
      0xb975d6b6ee39e436, 0xb3e2fd538e122b44,
      0xe7d34c64a9c85d44, 0x60dbbca87196b616,
      0x90e40fbeea1d3a4a, 0xbc8955e946fe31cd,
      0xb51d13aea4a488dd, 0x6babab6398bdbe41,
      0xe264589a4dcdab14, 0xc696963c7eed2dd1,
      0x8d7eb76070a08aec, 0xfc1e1de5cf543ca2,
      0xb0de65388cc8ada8, 0x3b25a55f43294bcb,
      0xdd15fe86affad912, 0x49ef0eb713f39ebe,
      0x8a2dbf142dfcc7ab, 0x6e3569326c784337,
      0xacb92ed9397bf996, 0x49c2c37f07965404,
      0xd7e77a8f87daf7fb, 0xdc33745ec97be906,
      0x86f0ac99b4e8dafd, 0x69a028bb3ded71a3,
      0xa8acd7c0222311bc, 0xc40832ea0d68ce0c,
      0xd2d80db02aabd62b, 0xf50a3fa490c30190,
      0x83c7088e1aab65db, 0x792667c6da79e0fa,
      0xa4b8cab1a1563f52, 0x577001b891185938,
      0xcde6fd5e09abcf26, 0xed4c0226b55e6f86,
      0x80b05e5ac60b6178, 0x544f8158315b05b4,
      0xa0dc75f1778e39d6, 0x696361ae3db1c721,
      0xc913936dd571c84c, 0x3bc3a19cd1e38e9,
      0xfb5878494ace3a5f, 0x4ab48a04065c723,
      0x9d174b2dcec0e47b, 0x62eb0d64283f9c76,
      0xc45d1df942711d9a, 0x3ba5d0bd324f8394,
      0xf5746577930d6500, 0xca8f44ec7ee36479,
      0x9968bf6abbe85f20, 0x7e998b13cf4e1ecb,
      0xbfc2ef456ae276e8, 0x9e3fedd8c321a67e,
      0xefb3ab16c59b14a2, 0xc5cfe94ef3ea101e,
      0x95d04aee3b80ece5, 0xbba1f1d158724a12,
      0xbb445da9ca61281f, 0x2a8a6e45ae8edc97,
      0xea1575143cf97226, 0xf52d09d71a3293bd,
      0x924d692ca61be758, 0x593c2626705f9c56,
      0xb6e0c377cfa2e12e, 0x6f8b2fb00c77836c,
      0xe498f455c38b997a, 0xb6dfb9c0f956447,
      0x8edf98b59a373fec, 0x4724bd4189bd5eac,
      0xb2977ee300c50fe7, 0x58edec91ec2cb657,
      0xdf3d5e9bc0f653e1, 0x2f2967b66737e3ed,
      0x8b865b215899f46c, 0xbd79e0d20082ee74,
      0xae67f1e9aec07187, 0xecd8590680a3aa11,
      0xda01ee641a708de9, 0xe80e6f4820cc9495,
      0x884134fe908658b2, 0x3109058d147fdcdd,
      0xaa51823e34a7eede, 0xbd4b46f0599fd415,
      0xd4e5e2cdc1d1ea96, 0x6c9e18ac7007c91a,
      0x850fadc09923329e, 0x3e2cf6bc604ddb0,
      0xa6539930bf6bff45, 0x84db8346b786151c,
      0xcfe87f7cef46ff16, 0xe612641865679a63,
      0x81f14fae158c5f6e, 0x4fcb7e8f3f60c07e,
      0xa26da3999aef7749, 0xe3be5e330f38f09d,
      0xcb090c8001ab551c, 0x5cadf5bfd3072cc5,
      0xfdcb4fa002162a63, 0x73d9732fc7c8f7f6,
      0x9e9f11c4014dda7e, 0x2867e7fddcdd9afa,
      0xc646d63501a1511d, 0xb281e1fd541501b8,
      0xf7d88bc24209a565, 0x1f225a7ca91a4226,
      0x9ae757596946075f, 0x3375788de9b06958,
      0xc1a12d2fc3978937, 0x52d6b1641c83ae,
      0xf209787bb47d6b84, 0xc0678c5dbd23a49a,
      0x9745eb4d50ce6332, 0xf840b7ba963646e0,
      0xbd176620a501fbff, 0xb650e5a93bc3d898,
      0xec5d3fa8ce427aff, 0xa3e51f138ab4cebe,
      0x93ba47c980e98cdf, 0xc66f336c36b10137,
      0xb8a8d9bbe123f017, 0xb80b0047445d4184,
      0xe6d3102ad96cec1d, 0xa60dc059157491e5,
      0x9043ea1ac7e41392, 0x87c89837ad68db2f,
      0xb454e4a179dd1877, 0x29babe4598c311fb,
      0xe16a1dc9d8545e94, 0xf4296dd6fef3d67a,
      0x8ce2529e2734bb1d, 0x1899e4a65f58660c,
      0xb01ae745b101e9e4, 0x5ec05dcff72e7f8f,
      0xdc21a1171d42645d, 0x76707543f4fa1f73,
      0x899504ae72497eba, 0x6a06494a791c53a8,
      0xabfa45da0edbde69, 0x487db9d17636892,
      0xd6f8d7509292d603, 0x45a9d2845d3c42b6,
      0x865b86925b9bc5c2, 0xb8a2392ba45a9b2,
      0xa7f26836f282b732, 0x8e6cac7768d7141e,
      0xd1ef0244af2364ff, 0x3207d795430cd926,
      0x8335616aed761f1f, 0x7f44e6bd49e807b8,
      0xa402b9c5a8d3a6e7, 0x5f16206c9c6209a6,
      0xcd036837130890a1, 0x36dba887c37a8c0f,
      0x802221226be55a64, 0xc2494954da2c9789,
      0xa02aa96b06deb0fd, 0xf2db9baa10b7bd6c,
      0xc83553c5c8965d3d, 0x6f92829494e5acc7,
      0xfa42a8b73abbf48c, 0xcb772339ba1f17f9,
      0x9c69a97284b578d7, 0xff2a760414536efb,
      0xc38413cf25e2d70d, 0xfef5138519684aba,
      0xf46518c2ef5b8cd1, 0x7eb258665fc25d69,
      0x98bf2f79d5993802, 0xef2f773ffbd97a61,
      0xbeeefb584aff8603, 0xaafb550ffacfd8fa,
      0xeeaaba2e5dbf6784, 0x95ba2a53f983cf38,
      0x952ab45cfa97a0b2, 0xdd945a747bf26183,
      0xba756174393d88df, 0x94f971119aeef9e4,
      0xe912b9d1478ceb17, 0x7a37cd5601aab85d,
      0x91abb422ccb812ee, 0xac62e055c10ab33a,
      0xb616a12b7fe617aa, 0x577b986b314d6009,
      0xe39c49765fdf9d94, 0xed5a7e85fda0b80b,
      0x8e41ade9fbebc27d, 0x14588f13be847307,
      0xb1d219647ae6b31c, 0x596eb2d8ae258fc8,
      0xde469fbd99a05fe3, 0x6fca5f8ed9aef3bb,
      0x8aec23d680043bee, 0x25de7bb9480d5854,
      0xada72ccc20054ae9, 0xaf561aa79a10ae6a,
      0xd910f7ff28069da4, 0x1b2ba1518094da04,
      0x87aa9aff79042286, 0x90fb44d2f05d0842,
      0xa99541bf57452b28, 0x353a1607ac744a53,
      0xd3fa922f2d1675f2, 0x42889b8997915ce8,
      0x847c9b5d7c2e09b7, 0x69956135febada11,
      0xa59bc234db398c25, 0x43fab9837e699095,
      0xcf02b2c21207ef2e, 0x94f967e45e03f4bb,
      0x8161afb94b44f57d, 0x1d1be0eebac278f5,
      0xa1ba1ba79e1632dc, 0x6462d92a69731732,
      0xca28a291859bbf93, 0x7d7b8f7503cfdcfe,
      0xfcb2cb35e702af78, 0x5cda735244c3d43e,
      0x9defbf01b061adab, 0x3a0888136afa64a7,
      0xc56baec21c7a1916, 0x88aaa1845b8fdd0,
      0xf6c69a72a3989f5b, 0x8aad549e57273d45,
      0x9a3c2087a63f6399, 0x36ac54e2f678864b,
      0xc0cb28a98fcf3c7f, 0x84576a1bb416a7dd,
      0xf0fdf2d3f3c30b9f, 0x656d44a2a11c51d5,
      0x969eb7c47859e743, 0x9f644ae5a4b1b325,
      0xbc4665b596706114, 0x873d5d9f0dde1fee,
      0xeb57ff22fc0c7959, 0xa90cb506d155a7ea,
      0x9316ff75dd87cbd8, 0x9a7f12442d588f2,
      0xb7dcbf5354e9bece, 0xc11ed6d538aeb2f,
      0xe5d3ef282a242e81, 0x8f1668c8a86da5fa,
      0x8fa475791a569d10, 0xf96e017d694487bc,
      0xb38d92d760ec4455, 0x37c981dcc395a9ac,
      0xe070f78d3927556a, 0x85bbe253f47b1417,
      0x8c469ab843b89562, 0x93956d7478ccec8e,
      0xaf58416654a6babb, 0x387ac8d1970027b2,
      0xdb2e51bfe9d0696a, 0x6997b05fcc0319e,
      0x88fcf317f22241e2, 0x441fece3bdf81f03,
      0xab3c2fddeeaad25a, 0xd527e81cad7626c3,
      0xd60b3bd56a5586f1, 0x8a71e223d8d3b074,
      0x85c7056562757456, 0xf6872d5667844e49,
      0xa738c6bebb12d16c, 0xb428f8ac016561db,
      0xd106f86e69d785c7, 0xe13336d701beba52,
      0x82a45b450226b39c, 0xecc0024661173473,
      0xa34d721642b06084, 0x27f002d7f95d0190,
      0xcc20ce9bd35c78a5, 0x31ec038df7b441f4,
      0xff290242c83396ce, 0x7e67047175a15271,
      0x9f79a169bd203e41, 0xf0062c6e984d386,
      0xc75809c42c684dd1, 0x52c07b78a3e60868,
      0xf92e0c3537826145, 0xa7709a56ccdf8a82,
      0x9bbcc7a142b17ccb, 0x88a66076400bb691,
      0xc2abf989935ddbfe, 0x6acff893d00ea435,
      0xf356f7ebf83552fe, 0x583f6b8c4124d43,
      0x98165af37b2153de, 0xc3727a337a8b704a,
      0xbe1bf1b059e9a8d6, 0x744f18c0592e4c5c,
      0xeda2ee1c7064130c, 0x1162def06f79df73,
      0x9485d4d1c63e8be7, 0x8addcb5645ac2ba8,
      0xb9a74a0637ce2ee1, 0x6d953e2bd7173692,
      0xe8111c87c5c1ba99, 0xc8fa8db6ccdd0437,
      0x910ab1d4db9914a0, 0x1d9c9892400a22a2,
      0xb54d5e4a127f59c8, 0x2503beb6d00cab4b,
      0xe2a0b5dc971f303a, 0x2e44ae64840fd61d,
      0x8da471a9de737e24, 0x5ceaecfed289e5d2,
      0xb10d8e1456105dad, 0x7425a83e872c5f47,
      0xdd50f1996b947518, 0xd12f124e28f77719,
      0x8a5296ffe33cc92f, 0x82bd6b70d99aaa6f,
      0xace73cbfdc0bfb7b, 0x636cc64d1001550b,
      0xd8210befd30efa5a, 0x3c47f7e05401aa4e,
      0x8714a775e3e95c78, 0x65acfaec34810a71,
      0xa8d9d1535ce3b396, 0x7f1839a741a14d0d,
      0xd31045a8341ca07c, 0x1ede48111209a050,
      0x83ea2b892091e44d, 0x934aed0aab460432,
      0xa4e4b66b68b65d60, 0xf81da84d5617853f,
      0xce1de40642e3f4b9, 0x36251260ab9d668e,
      0x80d2ae83e9ce78f3, 0xc1d72b7c6b426019,
      0xa1075a24e4421730, 0xb24cf65b8612f81f,
      0xc94930ae1d529cfc, 0xdee033f26797b627,
      0xfb9b7cd9a4a7443c, 0x169840ef017da3b1,
      0x9d412e0806e88aa5, 0x8e1f289560ee864e,
      0xc491798a08a2ad4e, 0xf1a6f2bab92a27e2,
      0xf5b5d7ec8acb58a2, 0xae10af696774b1db,
      0x9991a6f3d6bf1765, 0xacca6da1e0a8ef29,
      0xbff610b0cc6edd3f, 0x17fd090a58d32af3,
      0xeff394dcff8a948e, 0xddfc4b4cef07f5b0,
      0x95f83d0a1fb69cd9, 0x4abdaf101564f98e,
      0xbb764c4ca7a4440f, 0x9d6d1ad41abe37f1,
      0xea53df5fd18d5513, 0x84c86189216dc5ed,
      0x92746b9be2f8552c, 0x32fd3cf5b4e49bb4,
      0xb7118682dbb66a77, 0x3fbc8c33221dc2a1,
      0xe4d5e82392a40515, 0xfabaf3feaa5334a,
      0x8f05b1163ba6832d, 0x29cb4d87f2a7400e,
      0xb2c71d5bca9023f8, 0x743e20e9ef511012,
      0xdf78e4b2bd342cf6, 0x914da9246b255416,
      0x8bab8eefb6409c1a, 0x1ad089b6c2f7548e,
      0xae9672aba3d0c320, 0xa184ac2473b529b1,
      0xda3c0f568cc4f3e8, 0xc9e5d72d90a2741e,
      0x8865899617fb1871, 0x7e2fa67c7a658892,
      0xaa7eebfb9df9de8d, 0xddbb901b98feeab7,
      0xd51ea6fa85785631, 0x552a74227f3ea565,
      0x8533285c936b35de, 0xd53a88958f87275f,
      0xa67ff273b8460356, 0x8a892abaf368f137,
      0xd01fef10a657842c, 0x2d2b7569b0432d85,
      0x8213f56a67f6b29b, 0x9c3b29620e29fc73,
      0xa298f2c501f45f42, 0x8349f3ba91b47b8f,
      0xcb3f2f7642717713, 0x241c70a936219a73,
      0xfe0efb53d30dd4d7, 0xed238cd383aa0110,
      0x9ec95d1463e8a506, 0xf4363804324a40aa,
      0xc67bb4597ce2ce48, 0xb143c6053edcd0d5,
      0xf81aa16fdc1b81da, 0xdd94b7868e94050a,
      0x9b10a4e5e9913128, 0xca7cf2b4191c8326,
      0xc1d4ce1f63f57d72, 0xfd1c2f611f63a3f0,
      0xf24a01a73cf2dccf, 0xbc633b39673c8cec,
      0x976e41088617ca01, 0xd5be0503e085d813,
      0xbd49d14aa79dbc82, 0x4b2d8644d8a74e18,
      0xec9c459d51852ba2, 0xddf8e7d60ed1219e,
      0x93e1ab8252f33b45, 0xcabb90e5c942b503,
      0xb8da1662e7b00a17, 0x3d6a751f3b936243,
      0xe7109bfba19c0c9d, 0xcc512670a783ad4,
      0x906a617d450187e2, 0x27fb2b80668b24c5,
      0xb484f9dc9641e9da, 0xb1f9f660802dedf6,
      0xe1a63853bbd26451, 0x5e7873f8a0396973,
      0x8d07e33455637eb2, 0xdb0b487b6423e1e8,
      0xb049dc016abc5e5f, 0x91ce1a9a3d2cda62,
      0xdc5c5301c56b75f7, 0x7641a140cc7810fb,
      0x89b9b3e11b6329ba, 0xa9e904c87fcb0a9d,
      0xac2820d9623bf429, 0x546345fa9fbdcd44,
      0xd732290fbacaf133, 0xa97c177947ad4095,
      0x867f59a9d4bed6c0, 0x49ed8eabcccc485d,
      0xa81f301449ee8c70, 0x5c68f256bfff5a74,
      0xd226fc195c6a2f8c, 0x73832eec6fff3111,
      0x83585d8fd9c25db7, 0xc831fd53c5ff7eab,
      0xa42e74f3d032f525, 0xba3e7ca8b77f5e55,
      0xcd3a1230c43fb26f, 0x28ce1bd2e55f35eb,
      0x80444b5e7aa7cf85, 0x7980d163cf5b81b3,
      0xa0555e361951c366, 0xd7e105bcc332621f,
      0xc86ab5c39fa63440, 0x8dd9472bf3fefaa7,
      0xfa856334878fc150, 0xb14f98f6f0feb951,
      0x9c935e00d4b9d8d2, 0x6ed1bf9a569f33d3,
      0xc3b8358109e84f07, 0xa862f80ec4700c8,
      0xf4a642e14c6262c8, 0xcd27bb612758c0fa,
      0x98e7e9cccfbd7dbd, 0x8038d51cb897789c,
      0xbf21e44003acdd2c, 0xe0470a63e6bd56c3,
      0xeeea5d5004981478, 0x1858ccfce06cac74,
      0x95527a5202df0ccb, 0xf37801e0c43ebc8,
      0xbaa718e68396cffd, 0xd30560258f54e6ba,
      0xe950df20247c83fd, 0x47c6b82ef32a2069,
      0x91d28b7416cdd27e, 0x4cdc331d57fa5441,
      0xb6472e511c81471d, 0xe0133fe4adf8e952,
      0xe3d8f9e563a198e5, 0x58180fddd97723a6,
      0x8e679c2f5e44ff8f, 0x570f09eaa7ea7648,
  };
};

#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE

template <class unused>
constexpr uint64_t
    powers_template<unused>::power_of_five_128[number_of_entries];

#endif

using powers = powers_template<>;

} // namespace fast_float

#endif

#ifndef FASTFLOAT_DECIMAL_TO_BINARY_H
#define FASTFLOAT_DECIMAL_TO_BINARY_H

#include <cmath>
#include <cstdlib>
#include <cstring>

namespace fast_float {

// This will compute or rather approximate w * 5**q and return a pair of 64-bit
// words approximating the result, with the "high" part corresponding to the
// most significant bits and the low part corresponding to the least significant
// bits.
//
template <int bit_precision>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 value128
compute_product_approximation(int64_t q, uint64_t w) {
  int const index = 2 * int(q - powers::smallest_power_of_five);
  // For small values of q, e.g., q in [0,27], the answer is always exact
  // because The line value128 firstproduct = full_multiplication(w,
  // power_of_five_128[index]); gives the exact answer.
  value128 firstproduct =
      full_multiplication(w, powers::power_of_five_128[index]);
  static_assert((bit_precision >= 0) && (bit_precision <= 64),
                " precision should  be in (0,64]");
  constexpr uint64_t precision_mask =
      (bit_precision < 64) ? (uint64_t(0xFFFFFFFFFFFFFFFF) >> bit_precision)
                           : uint64_t(0xFFFFFFFFFFFFFFFF);
  if ((firstproduct.high & precision_mask) ==
      precision_mask) { // could further guard with  (lower + w < lower)
    // regarding the second product, we only need secondproduct.high, but our
    // expectation is that the compiler will optimize this extra work away if
    // needed.
    value128 secondproduct =
        full_multiplication(w, powers::power_of_five_128[index + 1]);
    firstproduct.low += secondproduct.high;
    if (secondproduct.high > firstproduct.low) {
      firstproduct.high++;
    }
  }
  return firstproduct;
}

namespace detail {
/**
 * For q in (0,350), we have that
 *  f = (((152170 + 65536) * q ) >> 16);
 * is equal to
 *   floor(p) + q
 * where
 *   p = log(5**q)/log(2) = q * log(5)/log(2)
 *
 * For negative values of q in (-400,0), we have that
 *  f = (((152170 + 65536) * q ) >> 16);
 * is equal to
 *   -ceil(p) + q
 * where
 *   p = log(5**-q)/log(2) = -q * log(5)/log(2)
 */
constexpr fastfloat_really_inline int32_t power(int32_t q) noexcept {
  return (((152170 + 65536) * q) >> 16) + 63;
}
} // namespace detail

// create an adjusted mantissa, biased by the invalid power2
// for significant digits already multiplied by 10 ** q.
template <typename binary>
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 adjusted_mantissa
compute_error_scaled(int64_t q, uint64_t w, int lz) noexcept {
  int hilz = int(w >> 63) ^ 1;
  adjusted_mantissa answer;
  answer.mantissa = w << hilz;
  int bias = binary::mantissa_explicit_bits() - binary::minimum_exponent();
  answer.power2 = int32_t(detail::power(int32_t(q)) + bias - hilz - lz - 62 +
                          invalid_am_bias);
  return answer;
}

// w * 10 ** q, without rounding the representation up.
// the power2 in the exponent will be adjusted by invalid_am_bias.
template <typename binary>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
compute_error(int64_t q, uint64_t w) noexcept {
  int lz = leading_zeroes(w);
  w <<= lz;
  value128 product =
      compute_product_approximation<binary::mantissa_explicit_bits() + 3>(q, w);
  return compute_error_scaled<binary>(q, product.high, lz);
}

// Computers w * 10 ** q.
// The returned value should be a valid number that simply needs to be
// packed. However, in some very rare cases, the computation will fail. In such
// cases, we return an adjusted_mantissa with a negative power of 2: the caller
// should recompute in such cases.
template <typename binary>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
compute_float(int64_t q, uint64_t w) noexcept {
  adjusted_mantissa answer;
  if ((w == 0) || (q < binary::smallest_power_of_ten())) {
    answer.power2 = 0;
    answer.mantissa = 0;
    // result should be zero
    return answer;
  }
  if (q > binary::largest_power_of_ten()) {
    // we want to get infinity:
    answer.power2 = binary::infinite_power();
    answer.mantissa = 0;
    return answer;
  }
  // At this point in time q is in [powers::smallest_power_of_five,
  // powers::largest_power_of_five].

  // We want the most significant bit of i to be 1. Shift if needed.
  int lz = leading_zeroes(w);
  w <<= lz;

  // The required precision is binary::mantissa_explicit_bits() + 3 because
  // 1. We need the implicit bit
  // 2. We need an extra bit for rounding purposes
  // 3. We might lose a bit due to the "upperbit" routine (result too small,
  // requiring a shift)

  value128 product =
      compute_product_approximation<binary::mantissa_explicit_bits() + 3>(q, w);
  // The computed 'product' is always sufficient.
  // Mathematical proof:
  // Noble Mushtak and Daniel Lemire, Fast Number Parsing Without Fallback (to
  // appear) See script/mushtak_lemire.py

  // The "compute_product_approximation" function can be slightly slower than a
  // branchless approach: value128 product = compute_product(q, w); but in
  // practice, we can win big with the compute_product_approximation if its
  // additional branch is easily predicted. Which is best is data specific.
  int upperbit = int(product.high >> 63);
  int shift = upperbit + 64 - binary::mantissa_explicit_bits() - 3;

  answer.mantissa = product.high >> shift;

  answer.power2 = int32_t(detail::power(int32_t(q)) + upperbit - lz -
                          binary::minimum_exponent());
  if (answer.power2 <= 0) { // we have a subnormal?
    // Here have that answer.power2 <= 0 so -answer.power2 >= 0
    if (-answer.power2 + 1 >=
        64) { // if we have more than 64 bits below the minimum exponent, you
              // have a zero for sure.
      answer.power2 = 0;
      answer.mantissa = 0;
      // result should be zero
      return answer;
    }
    // next line is safe because -answer.power2 + 1 < 64
    answer.mantissa >>= -answer.power2 + 1;
    // Thankfully, we can't have both "round-to-even" and subnormals because
    // "round-to-even" only occurs for powers close to 0 in the 32-bit and
    // and 64-bit case (with no more than 19 digits).
    answer.mantissa += (answer.mantissa & 1); // round up
    answer.mantissa >>= 1;
    // There is a weird scenario where we don't have a subnormal but just.
    // Suppose we start with 2.2250738585072013e-308, we end up
    // with 0x3fffffffffffff x 2^-1023-53 which is technically subnormal
    // whereas 0x40000000000000 x 2^-1023-53  is normal. Now, we need to round
    // up 0x3fffffffffffff x 2^-1023-53  and once we do, we are no longer
    // subnormal, but we can only know this after rounding.
    // So we only declare a subnormal if we are smaller than the threshold.
    answer.power2 =
        (answer.mantissa < (uint64_t(1) << binary::mantissa_explicit_bits()))
            ? 0
            : 1;
    return answer;
  }

  // usually, we round *up*, but if we fall right in between and and we have an
  // even basis, we need to round down
  // We are only concerned with the cases where 5**q fits in single 64-bit word.
  if ((product.low <= 1) && (q >= binary::min_exponent_round_to_even()) &&
      (q <= binary::max_exponent_round_to_even()) &&
      ((answer.mantissa & 3) == 1)) { // we may fall between two floats!
    // To be in-between two floats we need that in doing
    //   answer.mantissa = product.high >> (upperbit + 64 -
    //   binary::mantissa_explicit_bits() - 3);
    // ... we dropped out only zeroes. But if this happened, then we can go
    // back!!!
    if ((answer.mantissa << shift) == product.high) {
      answer.mantissa &= ~uint64_t(1); // flip it so that we do not round up
    }
  }

  answer.mantissa += (answer.mantissa & 1); // round up
  answer.mantissa >>= 1;
  if (answer.mantissa >= (uint64_t(2) << binary::mantissa_explicit_bits())) {
    answer.mantissa = (uint64_t(1) << binary::mantissa_explicit_bits());
    answer.power2++; // undo previous addition
  }

  answer.mantissa &= ~(uint64_t(1) << binary::mantissa_explicit_bits());
  if (answer.power2 >= binary::infinite_power()) { // infinity
    answer.power2 = binary::infinite_power();
    answer.mantissa = 0;
  }
  return answer;
}

} // namespace fast_float

#endif

#ifndef FASTFLOAT_BIGINT_H
#define FASTFLOAT_BIGINT_H

#include <cstring>


namespace fast_float {

// the limb width: we want efficient multiplication of double the bits in
// limb, or for 64-bit limbs, at least 64-bit multiplication where we can
// extract the high and low parts efficiently. this is every 64-bit
// architecture except for sparc, which emulates 128-bit multiplication.
// we might have platforms where `CHAR_BIT` is not 8, so let's avoid
// doing `8 * sizeof(limb)`.
#if defined(FASTFLOAT_64BIT) && !defined(__sparc)
#define FASTFLOAT_64BIT_LIMB 1
typedef uint64_t limb;
constexpr size_t limb_bits = 64;
#else
#define FASTFLOAT_32BIT_LIMB
typedef uint32_t limb;
constexpr size_t limb_bits = 32;
#endif

typedef span<limb> limb_span;

// number of bits in a bigint. this needs to be at least the number
// of bits required to store the largest bigint, which is
// `log2(10**(digits + max_exp))`, or `log2(10**(767 + 342))`, or
// ~3600 bits, so we round to 4000.
constexpr size_t bigint_bits = 4000;
constexpr size_t bigint_limbs = bigint_bits / limb_bits;

// vector-like type that is allocated on the stack. the entire
// buffer is pre-allocated, and only the length changes.
template <uint16_t size> struct stackvec {
  limb data[size];
  // we never need more than 150 limbs
  uint16_t length{0};

  stackvec() = default;
  stackvec(stackvec const &) = delete;
  stackvec &operator=(stackvec const &) = delete;
  stackvec(stackvec &&) = delete;
  stackvec &operator=(stackvec &&other) = delete;

  // create stack vector from existing limb span.
  FASTFLOAT_CONSTEXPR20 stackvec(limb_span s) {
    FASTFLOAT_ASSERT(try_extend(s));
  }

  FASTFLOAT_CONSTEXPR14 limb &operator[](size_t index) noexcept {
    FASTFLOAT_DEBUG_ASSERT(index < length);
    return data[index];
  }

  FASTFLOAT_CONSTEXPR14 const limb &operator[](size_t index) const noexcept {
    FASTFLOAT_DEBUG_ASSERT(index < length);
    return data[index];
  }

  // index from the end of the container
  FASTFLOAT_CONSTEXPR14 const limb &rindex(size_t index) const noexcept {
    FASTFLOAT_DEBUG_ASSERT(index < length);
    size_t rindex = length - index - 1;
    return data[rindex];
  }

  // set the length, without bounds checking.
  FASTFLOAT_CONSTEXPR14 void set_len(size_t len) noexcept {
    length = uint16_t(len);
  }

  constexpr size_t len() const noexcept { return length; }

  constexpr bool is_empty() const noexcept { return length == 0; }

  constexpr size_t capacity() const noexcept { return size; }

  // append item to vector, without bounds checking
  FASTFLOAT_CONSTEXPR14 void push_unchecked(limb value) noexcept {
    data[length] = value;
    length++;
  }

  // append item to vector, returning if item was added
  FASTFLOAT_CONSTEXPR14 bool try_push(limb value) noexcept {
    if (len() < capacity()) {
      push_unchecked(value);
      return true;
    } else {
      return false;
    }
  }

  // add items to the vector, from a span, without bounds checking
  FASTFLOAT_CONSTEXPR20 void extend_unchecked(limb_span s) noexcept {
    limb *ptr = data + length;
    tinyobj_ff::copy_n(s.ptr, s.len(), ptr);
    set_len(len() + s.len());
  }

  // try to add items to the vector, returning if items were added
  FASTFLOAT_CONSTEXPR20 bool try_extend(limb_span s) noexcept {
    if (len() + s.len() <= capacity()) {
      extend_unchecked(s);
      return true;
    } else {
      return false;
    }
  }

  // resize the vector, without bounds checking
  // if the new size is longer than the vector, assign value to each
  // appended item.
  FASTFLOAT_CONSTEXPR20
  void resize_unchecked(size_t new_len, limb value) noexcept {
    if (new_len > len()) {
      size_t count = new_len - len();
      limb *first = data + len();
      limb *last = first + count;
      tinyobj_ff::fill(first, last, value);
      set_len(new_len);
    } else {
      set_len(new_len);
    }
  }

  // try to resize the vector, returning if the vector was resized.
  FASTFLOAT_CONSTEXPR20 bool try_resize(size_t new_len, limb value) noexcept {
    if (new_len > capacity()) {
      return false;
    } else {
      resize_unchecked(new_len, value);
      return true;
    }
  }

  // check if any limbs are non-zero after the given index.
  // this needs to be done in reverse order, since the index
  // is relative to the most significant limbs.
  FASTFLOAT_CONSTEXPR14 bool nonzero(size_t index) const noexcept {
    while (index < len()) {
      if (rindex(index) != 0) {
        return true;
      }
      index++;
    }
    return false;
  }

  // normalize the big integer, so most-significant zero limbs are removed.
  FASTFLOAT_CONSTEXPR14 void normalize() noexcept {
    while (len() > 0 && rindex(0) == 0) {
      length--;
    }
  }
};

fastfloat_really_inline FASTFLOAT_CONSTEXPR14 uint64_t
empty_hi64(bool &truncated) noexcept {
  truncated = false;
  return 0;
}

fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t
uint64_hi64(uint64_t r0, bool &truncated) noexcept {
  truncated = false;
  int shl = leading_zeroes(r0);
  return r0 << shl;
}

fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t
uint64_hi64(uint64_t r0, uint64_t r1, bool &truncated) noexcept {
  int shl = leading_zeroes(r0);
  if (shl == 0) {
    truncated = r1 != 0;
    return r0;
  } else {
    int shr = 64 - shl;
    truncated = (r1 << shl) != 0;
    return (r0 << shl) | (r1 >> shr);
  }
}

fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t
uint32_hi64(uint32_t r0, bool &truncated) noexcept {
  return uint64_hi64(r0, truncated);
}

fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t
uint32_hi64(uint32_t r0, uint32_t r1, bool &truncated) noexcept {
  uint64_t x0 = r0;
  uint64_t x1 = r1;
  return uint64_hi64((x0 << 32) | x1, truncated);
}

fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t
uint32_hi64(uint32_t r0, uint32_t r1, uint32_t r2, bool &truncated) noexcept {
  uint64_t x0 = r0;
  uint64_t x1 = r1;
  uint64_t x2 = r2;
  return uint64_hi64(x0, (x1 << 32) | x2, truncated);
}

// add two small integers, checking for overflow.
// we want an efficient operation. for msvc, where
// we don't have built-in intrinsics, this is still
// pretty fast.
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 limb
scalar_add(limb x, limb y, bool &overflow) noexcept {
  limb z;
// gcc and clang
#if defined(__has_builtin)
#if __has_builtin(__builtin_add_overflow)
  if (!cpp20_and_in_constexpr()) {
    overflow = __builtin_add_overflow(x, y, &z);
    return z;
  }
#endif
#endif

  // generic, this still optimizes correctly on MSVC.
  z = x + y;
  overflow = z < x;
  return z;
}

// multiply two small integers, getting both the high and low bits.
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 limb
scalar_mul(limb x, limb y, limb &carry) noexcept {
#ifdef FASTFLOAT_64BIT_LIMB
#if defined(__SIZEOF_INT128__)
  // GCC and clang both define it as an extension.
  __uint128_t z = __uint128_t(x) * __uint128_t(y) + __uint128_t(carry);
  carry = limb(z >> limb_bits);
  return limb(z);
#else
  // fallback, no native 128-bit integer multiplication with carry.
  // on msvc, this optimizes identically, somehow.
  value128 z = full_multiplication(x, y);
  bool overflow;
  z.low = scalar_add(z.low, carry, overflow);
  z.high += uint64_t(overflow); // cannot overflow
  carry = z.high;
  return z.low;
#endif
#else
  uint64_t z = uint64_t(x) * uint64_t(y) + uint64_t(carry);
  carry = limb(z >> limb_bits);
  return limb(z);
#endif
}

// add scalar value to bigint starting from offset.
// used in grade school multiplication
template <uint16_t size>
inline FASTFLOAT_CONSTEXPR20 bool small_add_from(stackvec<size> &vec, limb y,
                                                 size_t start) noexcept {
  size_t index = start;
  limb carry = y;
  bool overflow;
  while (carry != 0 && index < vec.len()) {
    vec[index] = scalar_add(vec[index], carry, overflow);
    carry = limb(overflow);
    index += 1;
  }
  if (carry != 0) {
    FASTFLOAT_TRY(vec.try_push(carry));
  }
  return true;
}

// add scalar value to bigint.
template <uint16_t size>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool
small_add(stackvec<size> &vec, limb y) noexcept {
  return small_add_from(vec, y, 0);
}

// multiply bigint by scalar value.
template <uint16_t size>
inline FASTFLOAT_CONSTEXPR20 bool small_mul(stackvec<size> &vec,
                                            limb y) noexcept {
  limb carry = 0;
  for (size_t index = 0; index < vec.len(); index++) {
    vec[index] = scalar_mul(vec[index], y, carry);
  }
  if (carry != 0) {
    FASTFLOAT_TRY(vec.try_push(carry));
  }
  return true;
}

// add bigint to bigint starting from index.
// used in grade school multiplication
template <uint16_t size>
FASTFLOAT_CONSTEXPR20 bool large_add_from(stackvec<size> &x, limb_span y,
                                          size_t start) noexcept {
  // the effective x buffer is from `xstart..x.len()`, so exit early
  // if we can't get that current range.
  if (x.len() < start || y.len() > x.len() - start) {
    FASTFLOAT_TRY(x.try_resize(y.len() + start, 0));
  }

  bool carry = false;
  for (size_t index = 0; index < y.len(); index++) {
    limb xi = x[index + start];
    limb yi = y[index];
    bool c1 = false;
    bool c2 = false;
    xi = scalar_add(xi, yi, c1);
    if (carry) {
      xi = scalar_add(xi, 1, c2);
    }
    x[index + start] = xi;
    carry = c1 | c2;
  }

  // handle overflow
  if (carry) {
    FASTFLOAT_TRY(small_add_from(x, 1, y.len() + start));
  }
  return true;
}

// add bigint to bigint.
template <uint16_t size>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool
large_add_from(stackvec<size> &x, limb_span y) noexcept {
  return large_add_from(x, y, 0);
}

// grade-school multiplication algorithm
template <uint16_t size>
FASTFLOAT_CONSTEXPR20 bool long_mul(stackvec<size> &x, limb_span y) noexcept {
  limb_span xs = limb_span(x.data, x.len());
  stackvec<size> z(xs);
  limb_span zs = limb_span(z.data, z.len());

  if (y.len() != 0) {
    limb y0 = y[0];
    FASTFLOAT_TRY(small_mul(x, y0));
    for (size_t index = 1; index < y.len(); index++) {
      limb yi = y[index];
      stackvec<size> zi;
      if (yi != 0) {
        // re-use the same buffer throughout
        zi.set_len(0);
        FASTFLOAT_TRY(zi.try_extend(zs));
        FASTFLOAT_TRY(small_mul(zi, yi));
        limb_span zis = limb_span(zi.data, zi.len());
        FASTFLOAT_TRY(large_add_from(x, zis, index));
      }
    }
  }

  x.normalize();
  return true;
}

// grade-school multiplication algorithm
template <uint16_t size>
FASTFLOAT_CONSTEXPR20 bool large_mul(stackvec<size> &x, limb_span y) noexcept {
  if (y.len() == 1) {
    FASTFLOAT_TRY(small_mul(x, y[0]));
  } else {
    FASTFLOAT_TRY(long_mul(x, y));
  }
  return true;
}

template <typename = void> struct pow5_tables {
  static constexpr uint32_t large_step = 135;
  static constexpr uint64_t small_power_of_5[] = {
      1UL,
      5UL,
      25UL,
      125UL,
      625UL,
      3125UL,
      15625UL,
      78125UL,
      390625UL,
      1953125UL,
      9765625UL,
      48828125UL,
      244140625UL,
      1220703125UL,
      6103515625UL,
      30517578125UL,
      152587890625UL,
      762939453125UL,
      3814697265625UL,
      19073486328125UL,
      95367431640625UL,
      476837158203125UL,
      2384185791015625UL,
      11920928955078125UL,
      59604644775390625UL,
      298023223876953125UL,
      1490116119384765625UL,
      7450580596923828125UL,
  };
#ifdef FASTFLOAT_64BIT_LIMB
  constexpr static limb large_power_of_5[] = {
      1414648277510068013UL, 9180637584431281687UL, 4539964771860779200UL,
      10482974169319127550UL, 198276706040285095UL};
#else
  constexpr static limb large_power_of_5[] = {
      4279965485U, 329373468U,  4020270615U, 2137533757U, 4287402176U,
      1057042919U, 1071430142U, 2440757623U, 381945767U,  46164893U};
#endif
};

#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE

template <typename T> constexpr uint32_t pow5_tables<T>::large_step;

template <typename T> constexpr uint64_t pow5_tables<T>::small_power_of_5[];

template <typename T> constexpr limb pow5_tables<T>::large_power_of_5[];

#endif

// big integer type. implements a small subset of big integer
// arithmetic, using simple algorithms since asymptotically
// faster algorithms are slower for a small number of limbs.
// all operations assume the big-integer is normalized.
struct bigint : pow5_tables<> {
  // storage of the limbs, in little-endian order.
  stackvec<bigint_limbs> vec;

  FASTFLOAT_CONSTEXPR20 bigint() : vec() {}

  bigint(bigint const &) = delete;
  bigint &operator=(bigint const &) = delete;
  bigint(bigint &&) = delete;
  bigint &operator=(bigint &&other) = delete;

  FASTFLOAT_CONSTEXPR20 bigint(uint64_t value) : vec() {
#ifdef FASTFLOAT_64BIT_LIMB
    vec.push_unchecked(value);
#else
    vec.push_unchecked(uint32_t(value));
    vec.push_unchecked(uint32_t(value >> 32));
#endif
    vec.normalize();
  }

  // get the high 64 bits from the vector, and if bits were truncated.
  // this is to get the significant digits for the float.
  FASTFLOAT_CONSTEXPR20 uint64_t hi64(bool &truncated) const noexcept {
#ifdef FASTFLOAT_64BIT_LIMB
    if (vec.len() == 0) {
      return empty_hi64(truncated);
    } else if (vec.len() == 1) {
      return uint64_hi64(vec.rindex(0), truncated);
    } else {
      uint64_t result = uint64_hi64(vec.rindex(0), vec.rindex(1), truncated);
      truncated |= vec.nonzero(2);
      return result;
    }
#else
    if (vec.len() == 0) {
      return empty_hi64(truncated);
    } else if (vec.len() == 1) {
      return uint32_hi64(vec.rindex(0), truncated);
    } else if (vec.len() == 2) {
      return uint32_hi64(vec.rindex(0), vec.rindex(1), truncated);
    } else {
      uint64_t result =
          uint32_hi64(vec.rindex(0), vec.rindex(1), vec.rindex(2), truncated);
      truncated |= vec.nonzero(3);
      return result;
    }
#endif
  }

  // compare two big integers, returning the large value.
  // assumes both are normalized. if the return value is
  // negative, other is larger, if the return value is
  // positive, this is larger, otherwise they are equal.
  // the limbs are stored in little-endian order, so we
  // must compare the limbs in ever order.
  FASTFLOAT_CONSTEXPR20 int compare(bigint const &other) const noexcept {
    if (vec.len() > other.vec.len()) {
      return 1;
    } else if (vec.len() < other.vec.len()) {
      return -1;
    } else {
      for (size_t index = vec.len(); index > 0; index--) {
        limb xi = vec[index - 1];
        limb yi = other.vec[index - 1];
        if (xi > yi) {
          return 1;
        } else if (xi < yi) {
          return -1;
        }
      }
      return 0;
    }
  }

  // shift left each limb n bits, carrying over to the new limb
  // returns true if we were able to shift all the digits.
  FASTFLOAT_CONSTEXPR20 bool shl_bits(size_t n) noexcept {
    // Internally, for each item, we shift left by n, and add the previous
    // right shifted limb-bits.
    // For example, we transform (for u8) shifted left 2, to:
    //      b10100100 b01000010
    //      b10 b10010001 b00001000
    FASTFLOAT_DEBUG_ASSERT(n != 0);
    FASTFLOAT_DEBUG_ASSERT(n < sizeof(limb) * 8);

    size_t shl = n;
    size_t shr = limb_bits - shl;
    limb prev = 0;
    for (size_t index = 0; index < vec.len(); index++) {
      limb xi = vec[index];
      vec[index] = (xi << shl) | (prev >> shr);
      prev = xi;
    }

    limb carry = prev >> shr;
    if (carry != 0) {
      return vec.try_push(carry);
    }
    return true;
  }

  // move the limbs left by `n` limbs.
  FASTFLOAT_CONSTEXPR20 bool shl_limbs(size_t n) noexcept {
    FASTFLOAT_DEBUG_ASSERT(n != 0);
    if (n + vec.len() > vec.capacity()) {
      return false;
    } else if (!vec.is_empty()) {
      // move limbs
      limb *dst = vec.data + n;
      limb const *src = vec.data;
      tinyobj_ff::copy_backward(src, src + vec.len(), dst + vec.len());
      // fill in empty limbs
      limb *first = vec.data;
      limb *last = first + n;
      tinyobj_ff::fill(first, last, 0);
      vec.set_len(n + vec.len());
      return true;
    } else {
      return true;
    }
  }

  // move the limbs left by `n` bits.
  FASTFLOAT_CONSTEXPR20 bool shl(size_t n) noexcept {
    size_t rem = n % limb_bits;
    size_t div = n / limb_bits;
    if (rem != 0) {
      FASTFLOAT_TRY(shl_bits(rem));
    }
    if (div != 0) {
      FASTFLOAT_TRY(shl_limbs(div));
    }
    return true;
  }

  // get the number of leading zeros in the bigint.
  FASTFLOAT_CONSTEXPR20 int ctlz() const noexcept {
    if (vec.is_empty()) {
      return 0;
    } else {
#ifdef FASTFLOAT_64BIT_LIMB
      return leading_zeroes(vec.rindex(0));
#else
      // no use defining a specialized leading_zeroes for a 32-bit type.
      uint64_t r0 = vec.rindex(0);
      return leading_zeroes(r0 << 32);
#endif
    }
  }

  // get the number of bits in the bigint.
  FASTFLOAT_CONSTEXPR20 int bit_length() const noexcept {
    int lz = ctlz();
    return int(limb_bits * vec.len()) - lz;
  }

  FASTFLOAT_CONSTEXPR20 bool mul(limb y) noexcept { return small_mul(vec, y); }

  FASTFLOAT_CONSTEXPR20 bool add(limb y) noexcept { return small_add(vec, y); }

  // multiply as if by 2 raised to a power.
  FASTFLOAT_CONSTEXPR20 bool pow2(uint32_t exp) noexcept { return shl(exp); }

  // multiply as if by 5 raised to a power.
  FASTFLOAT_CONSTEXPR20 bool pow5(uint32_t exp) noexcept {
    // multiply by a power of 5
    size_t large_length = sizeof(large_power_of_5) / sizeof(limb);
    limb_span large = limb_span(large_power_of_5, large_length);
    while (exp >= large_step) {
      FASTFLOAT_TRY(large_mul(vec, large));
      exp -= large_step;
    }
#ifdef FASTFLOAT_64BIT_LIMB
    uint32_t small_step = 27;
    limb max_native = 7450580596923828125UL;
#else
    uint32_t small_step = 13;
    limb max_native = 1220703125U;
#endif
    while (exp >= small_step) {
      FASTFLOAT_TRY(small_mul(vec, max_native));
      exp -= small_step;
    }
    if (exp != 0) {
      // Work around clang bug https://godbolt.org/z/zedh7rrhc
      // This is similar to https://github.com/llvm/llvm-project/issues/47746,
      // except the workaround described there don't work here
      FASTFLOAT_TRY(small_mul(
          vec, limb(((void)small_power_of_5[0], small_power_of_5[exp]))));
    }

    return true;
  }

  // multiply as if by 10 raised to a power.
  FASTFLOAT_CONSTEXPR20 bool pow10(uint32_t exp) noexcept {
    FASTFLOAT_TRY(pow5(exp));
    return pow2(exp);
  }
};

} // namespace fast_float

#endif

#ifndef FASTFLOAT_DIGIT_COMPARISON_H
#define FASTFLOAT_DIGIT_COMPARISON_H

#include <cstring>


namespace fast_float {

// 1e0 to 1e19
constexpr static uint64_t powers_of_ten_uint64[] = {1UL,
                                                    10UL,
                                                    100UL,
                                                    1000UL,
                                                    10000UL,
                                                    100000UL,
                                                    1000000UL,
                                                    10000000UL,
                                                    100000000UL,
                                                    1000000000UL,
                                                    10000000000UL,
                                                    100000000000UL,
                                                    1000000000000UL,
                                                    10000000000000UL,
                                                    100000000000000UL,
                                                    1000000000000000UL,
                                                    10000000000000000UL,
                                                    100000000000000000UL,
                                                    1000000000000000000UL,
                                                    10000000000000000000UL};

// calculate the exponent, in scientific notation, of the number.
// this algorithm is not even close to optimized, but it has no practical
// effect on performance: in order to have a faster algorithm, we'd need
// to slow down performance for faster algorithms, and this is still fast.
template <typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 int32_t
scientific_exponent(parsed_number_string_t<UC> &num) noexcept {
  uint64_t mantissa = num.mantissa;
  int32_t exponent = int32_t(num.exponent);
  while (mantissa >= 10000) {
    mantissa /= 10000;
    exponent += 4;
  }
  while (mantissa >= 100) {
    mantissa /= 100;
    exponent += 2;
  }
  while (mantissa >= 10) {
    mantissa /= 10;
    exponent += 1;
  }
  return exponent;
}

// this converts a native floating-point number to an extended-precision float.
template <typename T>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
to_extended(T value) noexcept {
  using equiv_uint = equiv_uint_t<T>;
  constexpr equiv_uint exponent_mask = binary_format<T>::exponent_mask();
  constexpr equiv_uint mantissa_mask = binary_format<T>::mantissa_mask();
  constexpr equiv_uint hidden_bit_mask = binary_format<T>::hidden_bit_mask();

  adjusted_mantissa am;
  int32_t bias = binary_format<T>::mantissa_explicit_bits() -
                 binary_format<T>::minimum_exponent();
  equiv_uint bits;
#if FASTFLOAT_HAS_BIT_CAST
  bits = std::bit_cast<equiv_uint>(value);
#else
  ::memcpy(&bits, &value, sizeof(T));
#endif
  if ((bits & exponent_mask) == 0) {
    // denormal
    am.power2 = 1 - bias;
    am.mantissa = bits & mantissa_mask;
  } else {
    // normal
    am.power2 = int32_t((bits & exponent_mask) >>
                        binary_format<T>::mantissa_explicit_bits());
    am.power2 -= bias;
    am.mantissa = (bits & mantissa_mask) | hidden_bit_mask;
  }

  return am;
}

// get the extended precision value of the halfway point between b and b+u.
// we are given a native float that represents b, so we need to adjust it
// halfway between b and b+u.
template <typename T>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
to_extended_halfway(T value) noexcept {
  adjusted_mantissa am = to_extended(value);
  am.mantissa <<= 1;
  am.mantissa += 1;
  am.power2 -= 1;
  return am;
}

// round an extended-precision float to the nearest machine float.
template <typename T, typename callback>
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 void round(adjusted_mantissa &am,
                                                         callback cb) noexcept {
  int32_t mantissa_shift = 64 - binary_format<T>::mantissa_explicit_bits() - 1;
  if (-am.power2 >= mantissa_shift) {
    // have a denormal float
    int32_t shift = -am.power2 + 1;
    cb(am, tinyobj_ff::min_val<int32_t>(shift, 64));
    // check for round-up: if rounding-nearest carried us to the hidden bit.
    am.power2 = (am.mantissa <
                 (uint64_t(1) << binary_format<T>::mantissa_explicit_bits()))
                    ? 0
                    : 1;
    return;
  }

  // have a normal float, use the default shift.
  cb(am, mantissa_shift);

  // check for carry
  if (am.mantissa >=
      (uint64_t(2) << binary_format<T>::mantissa_explicit_bits())) {
    am.mantissa = (uint64_t(1) << binary_format<T>::mantissa_explicit_bits());
    am.power2++;
  }

  // check for infinite: we could have carried to an infinite power
  am.mantissa &= ~(uint64_t(1) << binary_format<T>::mantissa_explicit_bits());
  if (am.power2 >= binary_format<T>::infinite_power()) {
    am.power2 = binary_format<T>::infinite_power();
    am.mantissa = 0;
  }
}

template <typename callback>
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 void
round_nearest_tie_even(adjusted_mantissa &am, int32_t shift,
                       callback cb) noexcept {
  uint64_t const mask = (shift == 64) ? UINT64_MAX : (uint64_t(1) << shift) - 1;
  uint64_t const halfway = (shift == 0) ? 0 : uint64_t(1) << (shift - 1);
  uint64_t truncated_bits = am.mantissa & mask;
  bool is_above = truncated_bits > halfway;
  bool is_halfway = truncated_bits == halfway;

  // shift digits into position
  if (shift == 64) {
    am.mantissa = 0;
  } else {
    am.mantissa >>= shift;
  }
  am.power2 += shift;

  bool is_odd = (am.mantissa & 1) == 1;
  am.mantissa += uint64_t(cb(is_odd, is_halfway, is_above));
}

fastfloat_really_inline FASTFLOAT_CONSTEXPR14 void
round_down(adjusted_mantissa &am, int32_t shift) noexcept {
  if (shift == 64) {
    am.mantissa = 0;
  } else {
    am.mantissa >>= shift;
  }
  am.power2 += shift;
}

template <typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
skip_zeros(UC const *&first, UC const *last) noexcept {
  uint64_t val;
  while (!cpp20_and_in_constexpr() &&
         tinyobj_ff::distance(first, last) >= int_cmp_len<UC>()) {
    ::memcpy(&val, first, sizeof(uint64_t));
    if (val != int_cmp_zeros<UC>()) {
      break;
    }
    first += int_cmp_len<UC>();
  }
  while (first != last) {
    if (*first != UC('0')) {
      break;
    }
    first++;
  }
}

// determine if any non-zero digits were truncated.
// all characters must be valid digits.
template <typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool
is_truncated(UC const *first, UC const *last) noexcept {
  // do 8-bit optimizations, can just compare to 8 literal 0s.
  uint64_t val;
  while (!cpp20_and_in_constexpr() &&
         tinyobj_ff::distance(first, last) >= int_cmp_len<UC>()) {
    ::memcpy(&val, first, sizeof(uint64_t));
    if (val != int_cmp_zeros<UC>()) {
      return true;
    }
    first += int_cmp_len<UC>();
  }
  while (first != last) {
    if (*first != UC('0')) {
      return true;
    }
    ++first;
  }
  return false;
}

template <typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool
is_truncated(span<UC const> s) noexcept {
  return is_truncated(s.ptr, s.ptr + s.len());
}

template <typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
parse_eight_digits(UC const *&p, limb &value, size_t &counter,
                   size_t &count) noexcept {
  value = value * 100000000 + parse_eight_digits_unrolled(p);
  p += 8;
  counter += 8;
  count += 8;
}

template <typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 void
parse_one_digit(UC const *&p, limb &value, size_t &counter,
                size_t &count) noexcept {
  value = value * 10 + limb(*p - UC('0'));
  p++;
  counter++;
  count++;
}

fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
add_native(bigint &big, limb power, limb value) noexcept {
  big.mul(power);
  big.add(value);
}

fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
round_up_bigint(bigint &big, size_t &count) noexcept {
  // need to round-up the digits, but need to avoid rounding
  // ....9999 to ...10000, which could cause a false halfway point.
  add_native(big, 10, 1);
  count++;
}

// parse the significant digits into a big integer
template <typename UC>
inline FASTFLOAT_CONSTEXPR20 void
parse_mantissa(bigint &result, parsed_number_string_t<UC> &num,
               size_t max_digits, size_t &digits) noexcept {
  // try to minimize the number of big integer and scalar multiplication.
  // therefore, try to parse 8 digits at a time, and multiply by the largest
  // scalar value (9 or 19 digits) for each step.
  size_t counter = 0;
  digits = 0;
  limb value = 0;
#ifdef FASTFLOAT_64BIT_LIMB
  size_t step = 19;
#else
  size_t step = 9;
#endif

  // process all integer digits.
  UC const *p = num.integer.ptr;
  UC const *pend = p + num.integer.len();
  skip_zeros(p, pend);
  // process all digits, in increments of step per loop
  while (p != pend) {
    while ((tinyobj_ff::distance(p, pend) >= 8) && (step - counter >= 8) &&
           (max_digits - digits >= 8)) {
      parse_eight_digits(p, value, counter, digits);
    }
    while (counter < step && p != pend && digits < max_digits) {
      parse_one_digit(p, value, counter, digits);
    }
    if (digits == max_digits) {
      // add the temporary value, then check if we've truncated any digits
      add_native(result, limb(powers_of_ten_uint64[counter]), value);
      bool truncated = is_truncated(p, pend);
      if (num.fraction.ptr != nullptr) {
        truncated |= is_truncated(num.fraction);
      }
      if (truncated) {
        round_up_bigint(result, digits);
      }
      return;
    } else {
      add_native(result, limb(powers_of_ten_uint64[counter]), value);
      counter = 0;
      value = 0;
    }
  }

  // add our fraction digits, if they're available.
  if (num.fraction.ptr != nullptr) {
    p = num.fraction.ptr;
    pend = p + num.fraction.len();
    if (digits == 0) {
      skip_zeros(p, pend);
    }
    // process all digits, in increments of step per loop
    while (p != pend) {
      while ((tinyobj_ff::distance(p, pend) >= 8) && (step - counter >= 8) &&
             (max_digits - digits >= 8)) {
        parse_eight_digits(p, value, counter, digits);
      }
      while (counter < step && p != pend && digits < max_digits) {
        parse_one_digit(p, value, counter, digits);
      }
      if (digits == max_digits) {
        // add the temporary value, then check if we've truncated any digits
        add_native(result, limb(powers_of_ten_uint64[counter]), value);
        bool truncated = is_truncated(p, pend);
        if (truncated) {
          round_up_bigint(result, digits);
        }
        return;
      } else {
        add_native(result, limb(powers_of_ten_uint64[counter]), value);
        counter = 0;
        value = 0;
      }
    }
  }

  if (counter != 0) {
    add_native(result, limb(powers_of_ten_uint64[counter]), value);
  }
}

template <typename T>
inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
positive_digit_comp(bigint &bigmant, int32_t exponent) noexcept {
  FASTFLOAT_ASSERT(bigmant.pow10(uint32_t(exponent)));
  adjusted_mantissa answer;
  bool truncated;
  answer.mantissa = bigmant.hi64(truncated);
  int bias = binary_format<T>::mantissa_explicit_bits() -
             binary_format<T>::minimum_exponent();
  answer.power2 = bigmant.bit_length() - 64 + bias;

  round<T>(answer, [truncated](adjusted_mantissa &a, int32_t shift) {
    round_nearest_tie_even(
        a, shift,
        [truncated](bool is_odd, bool is_halfway, bool is_above) -> bool {
          return is_above || (is_halfway && truncated) ||
                 (is_odd && is_halfway);
        });
  });

  return answer;
}

// the scaling here is quite simple: we have, for the real digits `m * 10^e`,
// and for the theoretical digits `n * 2^f`. Since `e` is always negative,
// to scale them identically, we do `n * 2^f * 5^-f`, so we now have `m * 2^e`.
// we then need to scale by `2^(f- e)`, and then the two significant digits
// are of the same magnitude.
template <typename T>
inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa negative_digit_comp(
    bigint &bigmant, adjusted_mantissa am, int32_t exponent) noexcept {
  bigint &real_digits = bigmant;
  int32_t real_exp = exponent;

  // get the value of `b`, rounded down, and get a bigint representation of b+h
  adjusted_mantissa am_b = am;
  // gcc7 buf: use a lambda to remove the noexcept qualifier bug with
  // -Wnoexcept-type.
  round<T>(am_b,
           [](adjusted_mantissa &a, int32_t shift) { round_down(a, shift); });
  T b;
  to_float(false, am_b, b);
  adjusted_mantissa theor = to_extended_halfway(b);
  bigint theor_digits(theor.mantissa);
  int32_t theor_exp = theor.power2;

  // scale real digits and theor digits to be same power.
  int32_t pow2_exp = theor_exp - real_exp;
  uint32_t pow5_exp = uint32_t(-real_exp);
  if (pow5_exp != 0) {
    FASTFLOAT_ASSERT(theor_digits.pow5(pow5_exp));
  }
  if (pow2_exp > 0) {
    FASTFLOAT_ASSERT(theor_digits.pow2(uint32_t(pow2_exp)));
  } else if (pow2_exp < 0) {
    FASTFLOAT_ASSERT(real_digits.pow2(uint32_t(-pow2_exp)));
  }

  // compare digits, and use it to director rounding
  int ord = real_digits.compare(theor_digits);
  adjusted_mantissa answer = am;
  round<T>(answer, [ord](adjusted_mantissa &a, int32_t shift) {
    round_nearest_tie_even(
        a, shift, [ord](bool is_odd, bool _, bool __) -> bool {
          (void)_;  // not needed, since we've done our comparison
          (void)__; // not needed, since we've done our comparison
          if (ord > 0) {
            return true;
          } else if (ord < 0) {
            return false;
          } else {
            return is_odd;
          }
        });
  });

  return answer;
}

// parse the significant digits as a big integer to unambiguously round the
// the significant digits. here, we are trying to determine how to round
// an extended float representation close to `b+h`, halfway between `b`
// (the float rounded-down) and `b+u`, the next positive float. this
// algorithm is always correct, and uses one of two approaches. when
// the exponent is positive relative to the significant digits (such as
// 1234), we create a big-integer representation, get the high 64-bits,
// determine if any lower bits are truncated, and use that to direct
// rounding. in case of a negative exponent relative to the significant
// digits (such as 1.2345), we create a theoretical representation of
// `b` as a big-integer type, scaled to the same binary exponent as
// the actual digits. we then compare the big integer representations
// of both, and use that to direct rounding.
template <typename T, typename UC>
inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
digit_comp(parsed_number_string_t<UC> &num, adjusted_mantissa am) noexcept {
  // remove the invalid exponent bias
  am.power2 -= invalid_am_bias;

  int32_t sci_exp = scientific_exponent(num);
  size_t max_digits = binary_format<T>::max_digits();
  size_t digits = 0;
  bigint bigmant;
  parse_mantissa(bigmant, num, max_digits, digits);
  // can't underflow, since digits is at most max_digits.
  int32_t exponent = sci_exp + 1 - int32_t(digits);
  if (exponent >= 0) {
    return positive_digit_comp<T>(bigmant, exponent);
  } else {
    return negative_digit_comp<T>(bigmant, am, exponent);
  }
}

} // namespace fast_float

#endif

#ifndef FASTFLOAT_PARSE_NUMBER_H
#define FASTFLOAT_PARSE_NUMBER_H


#include <cmath>
#include <cstring>
#include <limits>

namespace fast_float {

namespace detail {
/**
 * Special case +inf, -inf, nan, infinity, -infinity.
 * The case comparisons could be made much faster given that we know that the
 * strings a null-free and fixed.
 **/
template <typename T, typename UC>
from_chars_result_t<UC>
    FASTFLOAT_CONSTEXPR14 parse_infnan(UC const *first, UC const *last,
                                       T &value, chars_format fmt) noexcept {
  from_chars_result_t<UC> answer{};
  answer.ptr = first;
  answer.ec = tinyobj_ff::ff_errc(); // be optimistic
  // assume first < last, so dereference without checks;
  bool const minusSign = (*first == UC('-'));
  // C++17 20.19.3.(7.1) explicitly forbids '+' sign here
  if ((*first == UC('-')) ||
      (uint64_t(fmt & chars_format::allow_leading_plus) &&
       (*first == UC('+')))) {
    ++first;
  }
  if (last - first >= 3) {
    if (fastfloat_strncasecmp(first, str_const_nan<UC>(), 3)) {
      answer.ptr = (first += 3);
      value = minusSign ? -std::numeric_limits<T>::quiet_NaN()
                        : std::numeric_limits<T>::quiet_NaN();
      // Check for possible nan(n-char-seq-opt), C++17 20.19.3.7,
      // C11 7.20.1.3.3. At least MSVC produces nan(ind) and nan(snan).
      if (first != last && *first == UC('(')) {
        for (UC const *ptr = first + 1; ptr != last; ++ptr) {
          if (*ptr == UC(')')) {
            answer.ptr = ptr + 1; // valid nan(n-char-seq-opt)
            break;
          } else if (!((UC('a') <= *ptr && *ptr <= UC('z')) ||
                       (UC('A') <= *ptr && *ptr <= UC('Z')) ||
                       (UC('0') <= *ptr && *ptr <= UC('9')) || *ptr == UC('_')))
            break; // forbidden char, not nan(n-char-seq-opt)
        }
      }
      return answer;
    }
    if (fastfloat_strncasecmp(first, str_const_inf<UC>(), 3)) {
      if ((last - first >= 8) &&
          fastfloat_strncasecmp(first + 3, str_const_inf<UC>() + 3, 5)) {
        answer.ptr = first + 8;
      } else {
        answer.ptr = first + 3;
      }
      value = minusSign ? -std::numeric_limits<T>::infinity()
                        : std::numeric_limits<T>::infinity();
      return answer;
    }
  }
  answer.ec = tinyobj_ff::ff_errc::invalid_argument;
  return answer;
}

/**
 * Returns true if the floating-pointing rounding mode is to 'nearest'.
 * It is the default on most system. This function is meant to be inexpensive.
 * Credit : @mwalcott3
 */
fastfloat_really_inline bool rounds_to_nearest() noexcept {
  // https://lemire.me/blog/2020/06/26/gcc-not-nearest/
#if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0)
  return false;
#endif
  // See
  // A fast function to check your floating-point rounding mode
  // https://lemire.me/blog/2022/11/16/a-fast-function-to-check-your-floating-point-rounding-mode/
  //
  // This function is meant to be equivalent to :
  // prior: #include <cfenv>
  //  return fegetround() == FE_TONEAREST;
  // However, it is expected to be much faster than the fegetround()
  // function call.
  //
  // The volatile keyword prevents the compiler from computing the function
  // at compile-time.
  // There might be other ways to prevent compile-time optimizations (e.g.,
  // asm). The value does not need to be std::numeric_limits<float>::min(), any
  // small value so that 1 + x should round to 1 would do (after accounting for
  // excess precision, as in 387 instructions).
  static float volatile fmin = std::numeric_limits<float>::min();
  float fmini = fmin; // we copy it so that it gets loaded at most once.
//
// Explanation:
// Only when fegetround() == FE_TONEAREST do we have that
// fmin + 1.0f == 1.0f - fmin.
//
// FE_UPWARD:
//  fmin + 1.0f > 1
//  1.0f - fmin == 1
//
// FE_DOWNWARD or  FE_TOWARDZERO:
//  fmin + 1.0f == 1
//  1.0f - fmin < 1
//
// Note: This may fail to be accurate if fast-math has been
// enabled, as rounding conventions may not apply.
#ifdef FASTFLOAT_VISUAL_STUDIO
#pragma warning(push)
//  todo: is there a VS warning?
//  see
//  https://stackoverflow.com/questions/46079446/is-there-a-warning-for-floating-point-equality-checking-in-visual-studio-2013
#elif defined(__clang__)
#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Wfloat-equal"
#elif defined(__GNUC__)
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wfloat-equal"
#endif
  return (fmini + 1.0f == 1.0f - fmini);
#ifdef FASTFLOAT_VISUAL_STUDIO
#pragma warning(pop)
#elif defined(__clang__)
#pragma clang diagnostic pop
#elif defined(__GNUC__)
#pragma GCC diagnostic pop
#endif
}

} // namespace detail

template <typename T> struct from_chars_caller {
  template <typename UC>
  FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC>
  call(UC const *first, UC const *last, T &value,
       parse_options_t<UC> options) noexcept {
    return from_chars_advanced(first, last, value, options);
  }
};

#ifdef __STDCPP_FLOAT32_T__
template <> struct from_chars_caller<std::float32_t> {
  template <typename UC>
  FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC>
  call(UC const *first, UC const *last, std::float32_t &value,
       parse_options_t<UC> options) noexcept {
    // if std::float32_t is defined, and we are in C++23 mode; macro set for
    // float32; set value to float due to equivalence between float and
    // float32_t
    float val;
    auto ret = from_chars_advanced(first, last, val, options);
    value = val;
    return ret;
  }
};
#endif

#ifdef __STDCPP_FLOAT64_T__
template <> struct from_chars_caller<std::float64_t> {
  template <typename UC>
  FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC>
  call(UC const *first, UC const *last, std::float64_t &value,
       parse_options_t<UC> options) noexcept {
    // if std::float64_t is defined, and we are in C++23 mode; macro set for
    // float64; set value as double due to equivalence between double and
    // float64_t
    double val;
    auto ret = from_chars_advanced(first, last, val, options);
    value = val;
    return ret;
  }
};
#endif

template <typename T, typename UC, typename>
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
from_chars(UC const *first, UC const *last, T &value,
           chars_format fmt /*= chars_format::general*/) noexcept {
  return from_chars_caller<T>::call(first, last, value,
                                    parse_options_t<UC>(fmt));
}

/**
 * This function overload takes parsed_number_string_t structure that is created
 * and populated either by from_chars_advanced function taking chars range and
 * parsing options or other parsing custom function implemented by user.
 */
template <typename T, typename UC>
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
from_chars_advanced(parsed_number_string_t<UC> &pns, T &value) noexcept {

  static_assert(is_supported_float_type<T>::value,
                "only some floating-point types are supported");
  static_assert(is_supported_char_type<UC>::value,
                "only char, wchar_t, char16_t and char32_t are supported");

  from_chars_result_t<UC> answer;

  answer.ec = tinyobj_ff::ff_errc(); // be optimistic
  answer.ptr = pns.lastmatch;
  // The implementation of the Clinger's fast path is convoluted because
  // we want round-to-nearest in all cases, irrespective of the rounding mode
  // selected on the thread.
  // We proceed optimistically, assuming that detail::rounds_to_nearest()
  // returns true.
  if (binary_format<T>::min_exponent_fast_path() <= pns.exponent &&
      pns.exponent <= binary_format<T>::max_exponent_fast_path() &&
      !pns.too_many_digits) {
    // Unfortunately, the conventional Clinger's fast path is only possible
    // when the system rounds to the nearest float.
    //
    // We expect the next branch to almost always be selected.
    // We could check it first (before the previous branch), but
    // there might be performance advantages at having the check
    // be last.
    if (!cpp20_and_in_constexpr() && detail::rounds_to_nearest()) {
      // We have that fegetround() == FE_TONEAREST.
      // Next is Clinger's fast path.
      if (pns.mantissa <= binary_format<T>::max_mantissa_fast_path()) {
        value = T(pns.mantissa);
        if (pns.exponent < 0) {
          value = value / binary_format<T>::exact_power_of_ten(-pns.exponent);
        } else {
          value = value * binary_format<T>::exact_power_of_ten(pns.exponent);
        }
        if (pns.negative) {
          value = -value;
        }
        return answer;
      }
    } else {
      // We do not have that fegetround() == FE_TONEAREST.
      // Next is a modified Clinger's fast path, inspired by Jakub Jelínek's
      // proposal
      if (pns.exponent >= 0 &&
          pns.mantissa <=
              binary_format<T>::max_mantissa_fast_path(pns.exponent)) {
#if defined(__clang__) || defined(FASTFLOAT_32BIT)
        // Clang may map 0 to -0.0 when fegetround() == FE_DOWNWARD
        if (pns.mantissa == 0) {
          value = pns.negative ? T(-0.) : T(0.);
          return answer;
        }
#endif
        value = T(pns.mantissa) *
                binary_format<T>::exact_power_of_ten(pns.exponent);
        if (pns.negative) {
          value = -value;
        }
        return answer;
      }
    }
  }
  adjusted_mantissa am =
      compute_float<binary_format<T>>(pns.exponent, pns.mantissa);
  if (pns.too_many_digits && am.power2 >= 0) {
    if (am != compute_float<binary_format<T>>(pns.exponent, pns.mantissa + 1)) {
      am = compute_error<binary_format<T>>(pns.exponent, pns.mantissa);
    }
  }
  // If we called compute_float<binary_format<T>>(pns.exponent, pns.mantissa)
  // and we have an invalid power (am.power2 < 0), then we need to go the long
  // way around again. This is very uncommon.
  if (am.power2 < 0) {
    am = digit_comp<T>(pns, am);
  }
  to_float(pns.negative, am, value);
  // Test for over/underflow.
  if ((pns.mantissa != 0 && am.mantissa == 0 && am.power2 == 0) ||
      am.power2 == binary_format<T>::infinite_power()) {
    answer.ec = tinyobj_ff::ff_errc::result_out_of_range;
  }
  return answer;
}

template <typename T, typename UC>
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
from_chars_float_advanced(UC const *first, UC const *last, T &value,
                          parse_options_t<UC> options) noexcept {

  static_assert(is_supported_float_type<T>::value,
                "only some floating-point types are supported");
  static_assert(is_supported_char_type<UC>::value,
                "only char, wchar_t, char16_t and char32_t are supported");

  chars_format const fmt = detail::adjust_for_feature_macros(options.format);

  from_chars_result_t<UC> answer;
  if (uint64_t(fmt & chars_format::skip_white_space)) {
    while ((first != last) && fast_float::is_space(*first)) {
      first++;
    }
  }
  if (first == last) {
    answer.ec = tinyobj_ff::ff_errc::invalid_argument;
    answer.ptr = first;
    return answer;
  }
  parsed_number_string_t<UC> pns =
      uint64_t(fmt & detail::basic_json_fmt)
          ? parse_number_string<true, UC>(first, last, options)
          : parse_number_string<false, UC>(first, last, options);
  if (!pns.valid) {
    if (uint64_t(fmt & chars_format::no_infnan)) {
      answer.ec = tinyobj_ff::ff_errc::invalid_argument;
      answer.ptr = first;
      return answer;
    } else {
      return detail::parse_infnan(first, last, value, fmt);
    }
  }

  // call overload that takes parsed_number_string_t directly.
  return from_chars_advanced(pns, value);
}

template <typename T, typename UC, typename>
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
from_chars(UC const *first, UC const *last, T &value, int base) noexcept {

  static_assert(is_supported_integer_type<T>::value,
                "only integer types are supported");
  static_assert(is_supported_char_type<UC>::value,
                "only char, wchar_t, char16_t and char32_t are supported");

  parse_options_t<UC> options;
  options.base = base;
  return from_chars_advanced(first, last, value, options);
}

template <typename T, typename UC>
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
from_chars_int_advanced(UC const *first, UC const *last, T &value,
                        parse_options_t<UC> options) noexcept {

  static_assert(is_supported_integer_type<T>::value,
                "only integer types are supported");
  static_assert(is_supported_char_type<UC>::value,
                "only char, wchar_t, char16_t and char32_t are supported");

  chars_format const fmt = detail::adjust_for_feature_macros(options.format);
  int const base = options.base;

  from_chars_result_t<UC> answer;
  if (uint64_t(fmt & chars_format::skip_white_space)) {
    while ((first != last) && fast_float::is_space(*first)) {
      first++;
    }
  }
  if (first == last || base < 2 || base > 36) {
    answer.ec = tinyobj_ff::ff_errc::invalid_argument;
    answer.ptr = first;
    return answer;
  }

  return parse_int_string(first, last, value, options);
}

template <size_t TypeIx> struct from_chars_advanced_caller {
  static_assert(TypeIx > 0, "unsupported type");
};

template <> struct from_chars_advanced_caller<1> {
  template <typename T, typename UC>
  FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC>
  call(UC const *first, UC const *last, T &value,
       parse_options_t<UC> options) noexcept {
    return from_chars_float_advanced(first, last, value, options);
  }
};

template <> struct from_chars_advanced_caller<2> {
  template <typename T, typename UC>
  FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC>
  call(UC const *first, UC const *last, T &value,
       parse_options_t<UC> options) noexcept {
    return from_chars_int_advanced(first, last, value, options);
  }
};

template <typename T, typename UC>
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
from_chars_advanced(UC const *first, UC const *last, T &value,
                    parse_options_t<UC> options) noexcept {
  return from_chars_advanced_caller<
      size_t(is_supported_float_type<T>::value) +
      2 * size_t(is_supported_integer_type<T>::value)>::call(first, last, value,
                                                             options);
}

} // namespace fast_float

#endif


// --- End embedded fast_float ---

// Clean up fast_float macros to avoid polluting the user's namespace.
#undef FASTFLOAT_32BIT
#undef FASTFLOAT_32BIT_LIMB
#undef FASTFLOAT_64BIT
#undef FASTFLOAT_64BIT_LIMB
#undef FASTFLOAT_ASCII_NUMBER_H
#undef FASTFLOAT_ASSERT
#undef FASTFLOAT_BIGINT_H
#undef FASTFLOAT_CONSTEXPR14
#undef FASTFLOAT_CONSTEXPR20
#undef FASTFLOAT_CONSTEXPR_FEATURE_DETECT_H
#undef FASTFLOAT_DEBUG_ASSERT
#undef FASTFLOAT_DECIMAL_TO_BINARY_H
#undef FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE
#undef FASTFLOAT_DIGIT_COMPARISON_H
#undef FASTFLOAT_ENABLE_IF
#undef FASTFLOAT_FAST_FLOAT_H
#undef FASTFLOAT_FAST_TABLE_H
#undef FASTFLOAT_FLOAT_COMMON_H
#undef FASTFLOAT_HAS_BIT_CAST
#undef FASTFLOAT_HAS_IS_CONSTANT_EVALUATED
#undef FASTFLOAT_HAS_SIMD
#undef FASTFLOAT_IF_CONSTEXPR17
#undef FASTFLOAT_IS_BIG_ENDIAN
#undef FASTFLOAT_IS_CONSTEXPR
#undef FASTFLOAT_NEON
#undef FASTFLOAT_PARSE_NUMBER_H
#undef fastfloat_really_inline
#undef FASTFLOAT_SIMD_DISABLE_WARNINGS
#undef FASTFLOAT_SIMD_RESTORE_WARNINGS
#undef FASTFLOAT_SSE2
#undef FASTFLOAT_STRINGIZE
#undef FASTFLOAT_STRINGIZE_IMPL
#undef FASTFLOAT_TRY
#undef FASTFLOAT_VERSION
#undef FASTFLOAT_VERSION_MAJOR
#undef FASTFLOAT_VERSION_MINOR
#undef FASTFLOAT_VERSION_PATCH
#undef FASTFLOAT_VERSION_STR
#undef FASTFLOAT_VISUAL_STUDIO

#endif  // TINYOBJLOADER_DISABLE_FAST_FLOAT

namespace tinyobj {

MaterialReader::~MaterialReader() {}

// Byte-stream reader for bounds-checked text parsing.
// Replaces raw `const char*` token pointers with `(buf, len, idx)` triple.
// Every byte access is guarded by an EOF check.
class StreamReader {
 public:
// Maximum number of bytes StreamReader will buffer from std::istream.
// Define this macro to a larger value if your application needs to parse
// very large streamed OBJ/MTL content.
#ifndef TINYOBJLOADER_STREAM_READER_MAX_BYTES
#define TINYOBJLOADER_STREAM_READER_MAX_BYTES (size_t(256) * size_t(1024) * size_t(1024))
#endif

  StreamReader(const char *buf, size_t length)
      : buf_(buf), length_(length), idx_(0), line_num_(1), col_num_(1) {}

  // Non-copyable, non-movable: buf_ may point into owned_buf_.
  StreamReader(const StreamReader &) /* = delete */;
  StreamReader &operator=(const StreamReader &) /* = delete */;

  // Build from std::istream by reading all content into an internal buffer.
  explicit StreamReader(std::istream &is) : buf_(NULL), length_(0), idx_(0), line_num_(1), col_num_(1) {
    const size_t max_stream_bytes = TINYOBJLOADER_STREAM_READER_MAX_BYTES;
    std::streampos start_pos = is.tellg();
    bool can_seek = (start_pos != std::streampos(-1));
    if (can_seek) {
      is.seekg(0, std::ios::end);
      std::streampos end_pos = is.tellg();
      if (end_pos >= start_pos) {
        std::streamoff remaining_off = static_cast<std::streamoff>(end_pos - start_pos);
        is.seekg(start_pos);
        unsigned long long remaining_ull = static_cast<unsigned long long>(remaining_off);
        if (remaining_ull > static_cast<unsigned long long>((std::numeric_limits<size_t>::max)())) {
          std::stringstream ss;
          ss << "input stream too large for this platform (" << remaining_ull
             << " bytes exceeds size_t max " << (std::numeric_limits<size_t>::max)() << ")\n";
          push_error(ss.str());
          buf_ = "";
          length_ = 0;
          return;
        }
        size_t remaining_size = static_cast<size_t>(remaining_ull);
        if (remaining_size > max_stream_bytes) {
          std::stringstream ss;
          ss << "input stream too large (" << remaining_size
             << " bytes exceeds limit " << max_stream_bytes << " bytes)\n";
          push_error(ss.str());
          buf_ = "";
          length_ = 0;
          return;
        }
        owned_buf_.resize(remaining_size);
        if (remaining_size > 0) {
          is.read(&owned_buf_[0], static_cast<std::streamsize>(remaining_size));
        }
        size_t actually_read = static_cast<size_t>(is.gcount());
        owned_buf_.resize(actually_read);
      }
    }
    if (!can_seek || owned_buf_.empty()) {
      // Stream doesn't support seeking, or seek probing failed.
      if (can_seek) is.seekg(start_pos);
      is.clear();
      std::vector<char> content;
      char chunk[4096];
      size_t total_read = 0;
      while (is.good()) {
        is.read(chunk, static_cast<std::streamsize>(sizeof(chunk)));
        std::streamsize nread = is.gcount();
        if (nread <= 0) break;
        size_t n = static_cast<size_t>(nread);
        if (n > (max_stream_bytes - total_read)) {
          std::stringstream ss;
          ss << "input stream too large (exceeds limit " << max_stream_bytes
             << " bytes)\n";
          push_error(ss.str());
          owned_buf_.clear();
          buf_ = "";
          length_ = 0;
          return;
        }
        content.insert(content.end(), chunk, chunk + n);
        total_read += n;
      }
      owned_buf_.swap(content);
    }
    buf_ = owned_buf_.empty() ? "" : &owned_buf_[0];
    length_ = owned_buf_.size();
  }

  bool eof() const { return idx_ >= length_; }
  size_t tell() const { return idx_; }
  size_t size() const { return length_; }
  size_t line_num() const { return line_num_; }
  size_t col_num() const { return col_num_; }

  char peek() const {
    if (idx_ >= length_) return '\0';
    return buf_[idx_];
  }

  char get() {
    if (idx_ >= length_) return '\0';
    char c = buf_[idx_++];
    if (c == '\n') { line_num_++; col_num_ = 1; } else { col_num_++; }
    return c;
  }

  void advance(size_t n) {
    for (size_t i = 0; i < n && idx_ < length_; i++) {
      if (buf_[idx_] == '\n') { line_num_++; col_num_ = 1; } else { col_num_++; }
      idx_++;
    }
  }

  void skip_space() {
    while (idx_ < length_ && (buf_[idx_] == ' ' || buf_[idx_] == '\t')) {
      col_num_++;
      idx_++;
    }
  }

  void skip_space_and_cr() {
    while (idx_ < length_ && (buf_[idx_] == ' ' || buf_[idx_] == '\t' || buf_[idx_] == '\r')) {
      col_num_++;
      idx_++;
    }
  }

  void skip_line() {
    while (idx_ < length_) {
      char c = buf_[idx_];
      if (c == '\n') {
        idx_++;
        line_num_++;
        col_num_ = 1;
        return;
      }
      if (c == '\r') {
        idx_++;
        if (idx_ < length_ && buf_[idx_] == '\n') {
          idx_++;
        }
        line_num_++;
        col_num_ = 1;
        return;
      }
      col_num_++;
      idx_++;
    }
  }

  bool at_line_end() const {
    if (idx_ >= length_) return true;
    char c = buf_[idx_];
    return (c == '\n' || c == '\r' || c == '\0');
  }

  std::string read_line() {
    std::string result;
    while (idx_ < length_) {
      char c = buf_[idx_];
      if (c == '\n' || c == '\r') break;
      result += c;
      col_num_++;
      idx_++;
    }
    return result;
  }

  // Reads a whitespace-delimited token. Used by tests and as a general utility.
  std::string read_token() {
    skip_space();
    std::string result;
    while (idx_ < length_) {
      char c = buf_[idx_];
      if (c == ' ' || c == '\t' || c == '\r' || c == '\n' || c == '\0') break;
      result += c;
      col_num_++;
      idx_++;
    }
    return result;
  }

  bool match(const char *prefix, size_t len) const {
    if (idx_ >= length_ || len > length_ - idx_) return false;
    return (memcmp(buf_ + idx_, prefix, len) == 0);
  }

  bool char_at(size_t offset, char c) const {
    if (idx_ >= length_ || offset >= length_ - idx_) return false;
    return buf_[idx_ + offset] == c;
  }

  char peek_at(size_t offset) const {
    if (idx_ >= length_ || offset >= length_ - idx_) return '\0';
    return buf_[idx_ + offset];
  }

  const char *current_ptr() const {
    if (idx_ >= length_) return "";
    return buf_ + idx_;
  }

  size_t remaining() const {
    return (idx_ < length_) ? (length_ - idx_) : 0;
  }

  // Returns the full text of the current line (for diagnostic display).
  std::string current_line_text() const {
    // Scan backward to find line start
    size_t line_start = idx_;
    while (line_start > 0 && buf_[line_start - 1] != '\n' && buf_[line_start - 1] != '\r') {
      line_start--;
    }
    // Scan forward to find line end
    size_t line_end = idx_;
    while (line_end < length_ && buf_[line_end] != '\n' && buf_[line_end] != '\r') {
      line_end++;
    }
    return std::string(buf_ + line_start, line_end - line_start);
  }

  // Clang-style formatted error with file:line:col and caret.
  std::string format_error(const std::string &filename, const std::string &msg) const {
    std::stringstream line_ss, col_ss;
    line_ss << line_num_;
    col_ss << col_num_;
    std::string result;
    result += filename + ":" + line_ss.str() + ":" + col_ss.str() + ": error: " + msg + "\n";
    std::string line_text = current_line_text();
    result += line_text + "\n";
    // Build caret line preserving tab alignment
    std::string caret;
    size_t caret_pos = (col_num_ > 0) ? (col_num_ - 1) : 0;
    for (size_t i = 0; i < caret_pos && i < line_text.size(); i++) {
      caret += (line_text[i] == '\t') ? '\t' : ' ';
    }
    caret += "^";
    result += caret + "\n";
    return result;
  }

  std::string format_error(const std::string &msg) const {
    return format_error("<input>", msg);
  }

  // Error stack
  void push_error(const std::string &msg) {
    errors_.push_back(msg);
  }

  void push_formatted_error(const std::string &filename, const std::string &msg) {
    errors_.push_back(format_error(filename, msg));
  }

  bool has_errors() const { return !errors_.empty(); }

  std::string get_errors() const {
    std::string result;
    for (size_t i = 0; i < errors_.size(); i++) {
      result += errors_[i];
    }
    return result;
  }

  const std::vector<std::string> &error_stack() const { return errors_; }

  void clear_errors() { errors_.clear(); }

 private:
  const char *buf_;
  size_t length_;
  size_t idx_;
  size_t line_num_;
  size_t col_num_;
  std::vector<char> owned_buf_;
  std::vector<std::string> errors_;
};

#ifdef TINYOBJLOADER_USE_MMAP
// RAII wrapper for memory-mapped file I/O.
// Opens a file and maps it into memory; the mapping is released on destruction.
// For empty files, data is set to "" and is_mapped remains false so close()
// will not attempt to unmap a string literal.
struct MappedFile {
  const char *data;
  size_t size;
  bool is_mapped;  // true when data points to an actual mapped region
#if defined(_WIN32)
  HANDLE hFile;
  HANDLE hMapping;
#else
  void *mapped_ptr;
#endif

  MappedFile() : data(NULL), size(0), is_mapped(false)
#if defined(_WIN32)
    , hFile(INVALID_HANDLE_VALUE), hMapping(NULL)
#else
    , mapped_ptr(NULL)
#endif
  {}

  // Opens and maps the file. Returns true on success.
  bool open(const char *filepath) {
#if defined(_WIN32)
    std::wstring wfilepath = LongPathW(UTF8ToWchar(std::string(filepath)));
    hFile = CreateFileW(wfilepath.c_str(), GENERIC_READ, FILE_SHARE_READ, NULL,
                        OPEN_EXISTING, FILE_ATTRIBUTE_NORMAL, NULL);
    if (hFile == INVALID_HANDLE_VALUE) return false;
    LARGE_INTEGER fileSize;
    if (!GetFileSizeEx(hFile, &fileSize)) { close(); return false; }
    if (fileSize.QuadPart < 0) { close(); return false; }
    unsigned long long fsize = static_cast<unsigned long long>(fileSize.QuadPart);
    if (fsize > static_cast<unsigned long long>((std::numeric_limits<size_t>::max)())) {
      close();
      return false;
    }
    size = static_cast<size_t>(fsize);
    if (size == 0) { data = ""; return true; }  // valid but empty; is_mapped stays false
    hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
    if (hMapping == NULL) { close(); return false; }
    data = static_cast<const char *>(MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0));
    if (!data) { close(); return false; }
    is_mapped = true;
    return true;
#else
    int fd = ::open(filepath, O_RDONLY);
    if (fd == -1) return false;
    struct stat sb;
    if (fstat(fd, &sb) != 0) { ::close(fd); return false; }
    if (sb.st_size < 0) { ::close(fd); return false; }
    if (static_cast<unsigned long long>(sb.st_size) >
        static_cast<unsigned long long>((std::numeric_limits<size_t>::max)())) {
      ::close(fd);
      return false;
    }
    size = static_cast<size_t>(sb.st_size);
    if (size == 0) { ::close(fd); data = ""; return true; }  // valid but empty
    mapped_ptr = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
    ::close(fd);
    if (mapped_ptr == MAP_FAILED) { mapped_ptr = NULL; size = 0; return false; }
    data = static_cast<const char *>(mapped_ptr);
    is_mapped = true;
    return true;
#endif
  }

  void close() {
#if defined(_WIN32)
    if (is_mapped && data) { UnmapViewOfFile(data); }
    data = NULL;
    is_mapped = false;
    if (hMapping != NULL) { CloseHandle(hMapping); hMapping = NULL; }
    if (hFile != INVALID_HANDLE_VALUE) { CloseHandle(hFile); hFile = INVALID_HANDLE_VALUE; }
#else
    if (is_mapped && mapped_ptr && mapped_ptr != MAP_FAILED) { munmap(mapped_ptr, size); }
    mapped_ptr = NULL;
    data = NULL;
    is_mapped = false;
#endif
    size = 0;
  }

  ~MappedFile() { close(); }

 private:
  MappedFile(const MappedFile &);             // non-copyable
  MappedFile &operator=(const MappedFile &);  // non-copyable
};
#endif  // TINYOBJLOADER_USE_MMAP


struct vertex_index_t {
  int v_idx, vt_idx, vn_idx;
  vertex_index_t() : v_idx(-1), vt_idx(-1), vn_idx(-1) {}
  explicit vertex_index_t(int idx) : v_idx(idx), vt_idx(idx), vn_idx(idx) {}
  vertex_index_t(int vidx, int vtidx, int vnidx)
      : v_idx(vidx), vt_idx(vtidx), vn_idx(vnidx) {}
};

// Internal data structure for face representation
// index + smoothing group.
struct face_t {
  unsigned int
      smoothing_group_id;  // smoothing group id. 0 = smoothing groupd is off.
  int pad_;
  std::vector<vertex_index_t> vertex_indices;  // face vertex indices.

  face_t() : smoothing_group_id(0), pad_(0) {}
};

// Internal data structure for line representation
struct __line_t {
  // l v1/vt1 v2/vt2 ...
  // In the specification, line primitrive does not have normal index, but
  // TinyObjLoader allow it
  std::vector<vertex_index_t> vertex_indices;
};

// Internal data structure for points representation
struct __points_t {
  // p v1 v2 ...
  // In the specification, point primitrive does not have normal index and
  // texture coord index, but TinyObjLoader allow it.
  std::vector<vertex_index_t> vertex_indices;
};

struct tag_sizes {
  tag_sizes() : num_ints(0), num_reals(0), num_strings(0) {}
  int num_ints;
  int num_reals;
  int num_strings;
};

struct obj_shape {
  std::vector<real_t> v;
  std::vector<real_t> vn;
  std::vector<real_t> vt;
};

//
// Manages group of primitives(face, line, points, ...)
struct PrimGroup {
  std::vector<face_t> faceGroup;
  std::vector<__line_t> lineGroup;
  std::vector<__points_t> pointsGroup;

  void clear() {
    faceGroup.clear();
    lineGroup.clear();
    pointsGroup.clear();
  }

  bool IsEmpty() const {
    return faceGroup.empty() && lineGroup.empty() && pointsGroup.empty();
  }

  // TODO(syoyo): bspline, surface, ...
};

// See
// http://stackoverflow.com/questions/6089231/getting-std-ifstream-to-handle-lf-cr-and-crlf
#define IS_SPACE(x) (((x) == ' ') || ((x) == '\t'))
#define IS_DIGIT(x) \
  (static_cast<unsigned int>((x) - '0') < static_cast<unsigned int>(10))
#define IS_NEW_LINE(x) (((x) == '\r') || ((x) == '\n') || ((x) == '\0'))

template <typename T>
static inline std::string toString(const T &t) {
  std::stringstream ss;
  ss << t;
  return ss.str();
}

static inline std::string removeUtf8Bom(const std::string& input) {
    // UTF-8 BOM = 0xEF,0xBB,0xBF
    if (input.size() >= 3 &&
        static_cast<unsigned char>(input[0]) == 0xEF &&
        static_cast<unsigned char>(input[1]) == 0xBB &&
        static_cast<unsigned char>(input[2]) == 0xBF) {
        return input.substr(3); // Skip BOM
    }
    return input;
}

// Trim trailing spaces and tabs from a string.
static inline std::string trimTrailingWhitespace(const std::string &s) {
  size_t end = s.find_last_not_of(" \t");
  if (end == std::string::npos) return "";
  return s.substr(0, end + 1);
}

struct warning_context {
  std::string *warn;
  size_t line_number;
  std::string filename;
};

// Safely convert size_t to int, clamping at INT_MAX to prevent overflow.
static inline int size_to_int(size_t sz) {
  return sz > static_cast<size_t>(INT_MAX) ? INT_MAX : static_cast<int>(sz);
}

// Make index zero-base, and also support relative index.
static inline bool fixIndex(int idx, int n, int *ret, bool allow_zero,
                            const warning_context &context) {
  if (!ret) {
    return false;
  }

  if (idx > 0) {
    (*ret) = idx - 1;
    return true;
  }

  if (idx == 0) {
    // zero is not allowed according to the spec.
    if (context.warn) {
      (*context.warn) +=
          context.filename + ":" + toString(context.line_number) +
          ": warning: zero value index found (will have a value of -1 for "
          "normal and tex indices)\n";
    }

    (*ret) = idx - 1;
    return allow_zero;
  }

  if (idx < 0) {
    (*ret) = n + idx;  // negative value = relative
    if ((*ret) < 0) {
      return false;  // invalid relative index
    }
    return true;
  }

  return false;  // never reach here.
}

static inline std::string parseString(const char **token) {
  std::string s;
  (*token) += strspn((*token), " \t");
  size_t e = strcspn((*token), " \t\r");
  s = std::string((*token), &(*token)[e]);
  (*token) += e;
  return s;
}

static inline int parseInt(const char **token) {
  (*token) += strspn((*token), " \t");
  int i = atoi((*token));
  (*token) += strcspn((*token), " \t\r");
  return i;
}

#ifndef TINYOBJLOADER_DISABLE_FAST_FLOAT

// ---- fast_float-based float parser (bit-exact with strtod, ~3x faster) ----

namespace detail_fp {

// Case-insensitive prefix match. Returns pointer past matched prefix, or NULL.
static inline const char *match_iprefix(const char *p, const char *end,
                                        const char *prefix) {
  while (*prefix) {
    if (p == end) return NULL;
    char c = *p;
    char e = *prefix;
    if (c >= 'A' && c <= 'Z') c += 32;
    if (e >= 'A' && e <= 'Z') e += 32;
    if (c != e) return NULL;
    ++p;
    ++prefix;
  }
  return p;
}

// Try to parse nan/inf. Returns true if matched, sets *result and *end_ptr.
static inline bool tryParseNanInf(const char *first, const char *last,
                                  double *result, const char **end_ptr) {
  if (first >= last) return false;

  const char *p = first;
  bool negative = false;

  if (*p == '-') {
    negative = true;
    ++p;
  } else if (*p == '+') {
    ++p;
  }

  if (p >= last) return false;

  // Try "nan"
  const char *after = match_iprefix(p, last, "nan");
  if (after) {
    *result = 0.0;  // nan -> 0.0 for OBJ
    *end_ptr = after;
    return true;
  }

  // Try "infinity" first (longer match), then "inf"
  after = match_iprefix(p, last, "infinity");
  if (after) {
    *result = negative ? std::numeric_limits<double>::lowest()
                       : (std::numeric_limits<double>::max)();
    *end_ptr = after;
    return true;
  }

  after = match_iprefix(p, last, "inf");
  if (after) {
    *result = negative ? std::numeric_limits<double>::lowest()
                       : (std::numeric_limits<double>::max)();
    *end_ptr = after;
    return true;
  }

  return false;
}

}  // namespace detail_fp

// Tries to parse a floating point number located at s.
// Uses fast_float::from_chars for bit-exact, high-performance parsing.
// Handles OBJ quirks: leading '+', nan/inf with replacement values.
//
// s_end should be a location in the string where reading should absolutely
// stop. For example at the end of the string, to prevent buffer overflows.
//
// If the parsing is a success, result is set to the parsed value and true
// is returned.
//
static bool tryParseDouble(const char *s, const char *s_end, double *result) {
  if (!s || !s_end || !result || s >= s_end) {
    return false;
  }

  // Check for nan/inf (starts with [nNiI] or [+-] followed by [nNiI])
  const char *p = s;
  if (p < s_end && (*p == '+' || *p == '-')) ++p;
  if (p < s_end) {
    char fc = *p;
    if (fc >= 'A' && fc <= 'Z') fc += 32;
    if (fc == 'n' || fc == 'i') {
      const char *end_ptr;
      if (detail_fp::tryParseNanInf(s, s_end, result, &end_ptr)) {
        return true;
      }
    }
  }

  // Use allow_leading_plus so fast_float handles '+' natively.
  double tmp;
  auto r = fast_float::from_chars(s, s_end, tmp,
      fast_float::chars_format::general |
      fast_float::chars_format::allow_leading_plus);
  if (r.ec == tinyobj_ff::ff_errc::ok) {
    *result = tmp;
    return true;
  }
  // On error (invalid_argument, result_out_of_range), *result is unchanged.

  return false;
}

static inline real_t parseReal(const char **token, double default_value = 0.0) {
  (*token) += strspn((*token), " \t");
  const char *end = (*token) + strcspn((*token), " \t\r");
  double val = default_value;
  tryParseDouble((*token), end, &val);
  real_t f = static_cast<real_t>(val);
  (*token) = end;
  return f;
}

static inline bool parseReal(const char **token, real_t *out) {
  (*token) += strspn((*token), " \t");
  const char *end = (*token) + strcspn((*token), " \t\r");
  double val;
  bool ret = tryParseDouble((*token), end, &val);
  if (ret) {
    real_t f = static_cast<real_t>(val);
    (*out) = f;
  }
  (*token) = end;
  return ret;
}

#else  // TINYOBJLOADER_DISABLE_FAST_FLOAT

// ---- Legacy hand-written float parser (fallback) ----

// Tries to parse a floating point number located at s.
//
// s_end should be a location in the string where reading should absolutely
// stop. For example at the end of the string, to prevent buffer overflows.
//
// Parses the following EBNF grammar:
//   sign    = "+" | "-" ;
//   END     = ? anything not in digit ?
//   digit   = "0" | "1" | "2" | "3" | "4" | "5" | "6" | "7" | "8" | "9" ;
//   integer = [sign] , digit , {digit} ;
//   decimal = integer , ["." , integer] ;
//   float   = ( decimal , END ) | ( decimal , ("E" | "e") , integer , END ) ;
//
//  Valid strings are for example:
//   -0  +3.1417e+2  -0.0E-3  1.0324  -1.41   11e2
//
// If the parsing is a success, result is set to the parsed value and true
// is returned.
//
// The function is greedy and will parse until any of the following happens:
//  - a non-conforming character is encountered.
//  - s_end is reached.
//
// The following situations triggers a failure:
//  - s >= s_end.
//  - parse failure.
//
static bool tryParseDouble(const char *s, const char *s_end, double *result) {
  if (s >= s_end) {
    return false;
  }

  double mantissa = 0.0;
  // This exponent is base 2 rather than 10.
  // However the exponent we parse is supposed to be one of ten,
  // thus we must take care to convert the exponent/and or the
  // mantissa to a * 2^E, where a is the mantissa and E is the
  // exponent.
  // To get the final double we will use ldexp, it requires the
  // exponent to be in base 2.
  int exponent = 0;

  // NOTE: THESE MUST BE DECLARED HERE SINCE WE ARE NOT ALLOWED
  // TO JUMP OVER DEFINITIONS.
  char sign = '+';
  char exp_sign = '+';
  char const *curr = s;

  // How many characters were read in a loop.
  int read = 0;
  // Tells whether a loop terminated due to reaching s_end.
  bool end_not_reached = false;
  bool leading_decimal_dots = false;

  /*
          BEGIN PARSING.
  */

  // Find out what sign we've got.
  if (*curr == '+' || *curr == '-') {
    sign = *curr;
    curr++;
    if ((curr != s_end) && (*curr == '.')) {
      // accept. Somethig like `.7e+2`, `-.5234`
      leading_decimal_dots = true;
    }
  } else if (IS_DIGIT(*curr)) { /* Pass through. */
  } else if (*curr == '.') {
    // accept. Somethig like `.7e+2`, `-.5234`
    leading_decimal_dots = true;
  } else {
    goto fail;
  }

  // Read the integer part.
  end_not_reached = (curr != s_end);
  if (!leading_decimal_dots) {
    while (end_not_reached && IS_DIGIT(*curr)) {
      mantissa *= 10;
      mantissa += static_cast<int>(*curr - 0x30);
      curr++;
      read++;
      end_not_reached = (curr != s_end);
    }

    // We must make sure we actually got something.
    if (read == 0) goto fail;
  }

  // We allow numbers of form "#", "###" etc.
  if (!end_not_reached) goto assemble;

  // Read the decimal part.
  if (*curr == '.') {
    curr++;
    read = 1;
    end_not_reached = (curr != s_end);
    while (end_not_reached && IS_DIGIT(*curr)) {
      static const double pow_lut[] = {
          1.0, 0.1, 0.01, 0.001, 0.0001, 0.00001, 0.000001, 0.0000001,
      };
      const int lut_entries = sizeof pow_lut / sizeof pow_lut[0];

      // NOTE: Don't use powf here, it will absolutely murder precision.
      mantissa += static_cast<int>(*curr - 0x30) *
                  (read < lut_entries ? pow_lut[read] : std::pow(10.0, -read));
      read++;
      curr++;
      end_not_reached = (curr != s_end);
    }
  } else if (*curr == 'e' || *curr == 'E') {
  } else {
    goto assemble;
  }

  if (!end_not_reached) goto assemble;

  // Read the exponent part.
  if (*curr == 'e' || *curr == 'E') {
    curr++;
    // Figure out if a sign is present and if it is.
    end_not_reached = (curr != s_end);
    if (end_not_reached && (*curr == '+' || *curr == '-')) {
      exp_sign = *curr;
      curr++;
    } else if (IS_DIGIT(*curr)) { /* Pass through. */
    } else {
      // Empty E is not allowed.
      goto fail;
    }

    read = 0;
    end_not_reached = (curr != s_end);
    while (end_not_reached && IS_DIGIT(*curr)) {
      // To avoid annoying MSVC's min/max macro definiton,
      // Use hardcoded int max value
      if (exponent >
          ((2147483647 - 9) / 10)) {  // (INT_MAX - 9) / 10, guards both multiply and add
        // Integer overflow
        goto fail;
      }
      exponent *= 10;
      exponent += static_cast<int>(*curr - 0x30);
      curr++;
      read++;
      end_not_reached = (curr != s_end);
    }
    exponent *= (exp_sign == '+' ? 1 : -1);
    if (read == 0) goto fail;
  }

assemble:
  *result = (sign == '+' ? 1 : -1) *
            (exponent ? std::ldexp(mantissa * std::pow(5.0, exponent), exponent)
                      : mantissa);
  return true;
fail:
  return false;
}

static inline real_t parseReal(const char **token, double default_value = 0.0) {
  (*token) += strspn((*token), " \t");
  const char *end = (*token) + strcspn((*token), " \t\r");
  double val = default_value;
  tryParseDouble((*token), end, &val);
  real_t f = static_cast<real_t>(val);
  (*token) = end;
  return f;
}

static inline bool parseReal(const char **token, real_t *out) {
  (*token) += strspn((*token), " \t");
  const char *end = (*token) + strcspn((*token), " \t\r");
  double val;
  bool ret = tryParseDouble((*token), end, &val);
  if (ret) {
    real_t f = static_cast<real_t>(val);
    (*out) = f;
  }
  (*token) = end;
  return ret;
}

#endif  // TINYOBJLOADER_DISABLE_FAST_FLOAT

static inline void parseReal2(real_t *x, real_t *y, const char **token,
                              const double default_x = 0.0,
                              const double default_y = 0.0) {
  (*x) = parseReal(token, default_x);
  (*y) = parseReal(token, default_y);
}

static inline void parseReal3(real_t *x, real_t *y, real_t *z,
                              const char **token, const double default_x = 0.0,
                              const double default_y = 0.0,
                              const double default_z = 0.0) {
  (*x) = parseReal(token, default_x);
  (*y) = parseReal(token, default_y);
  (*z) = parseReal(token, default_z);
}

#if 0  // not used
static inline void parseV(real_t *x, real_t *y, real_t *z, real_t *w,
                          const char **token, const double default_x = 0.0,
                          const double default_y = 0.0,
                          const double default_z = 0.0,
                          const double default_w = 1.0) {
  (*x) = parseReal(token, default_x);
  (*y) = parseReal(token, default_y);
  (*z) = parseReal(token, default_z);
  (*w) = parseReal(token, default_w);
}
#endif

// Extension: parse vertex with colors(6 items)
// Return 3: xyz, 4: xyzw, 6: xyzrgb
// `r`: red(case 6) or [w](case 4)
// NOTE: This classic path caps at 6 components per the de-facto vertex-color
// extension (xyz / xyzw / xyzrgb; weight and color are mutually exclusive) and
// ignores any further tokens.  LoadObjOpt / the experimental stream loader
// additionally accept a non-standard 7-component `v x y z w r g b` (weight +
// color) form, so they diverge from this function for 7+ token vertices.
static inline int parseVertexWithColor(real_t *x, real_t *y, real_t *z,
                                       real_t *r, real_t *g, real_t *b,
                                       const char **token,
                                       const double default_x = 0.0,
                                       const double default_y = 0.0,
                                       const double default_z = 0.0) {
  // TODO: Check error
  (*x) = parseReal(token, default_x);
  (*y) = parseReal(token, default_y);
  (*z) = parseReal(token, default_z);

  // - 4 components(x, y, z, w) ot 6 components
  bool has_r = parseReal(token, r);

  if (!has_r) {
    (*r) = (*g) = (*b) = 1.0;
    return 3;
  }

  bool has_g = parseReal(token, g);

  if (!has_g) {
    (*g) = (*b) = 1.0;
    return 4;
  }

  bool has_b = parseReal(token, b);

  if (!has_b) {
    (*r) = (*g) = (*b) = 1.0;
    return 3;  // treated as xyz
  }

  return 6;
}

static inline bool parseOnOff(const char **token, bool default_value = true) {
  (*token) += strspn((*token), " \t");
  const char *end = (*token) + strcspn((*token), " \t\r");

  bool ret = default_value;
  if ((0 == strncmp((*token), "on", 2))) {
    ret = true;
  } else if ((0 == strncmp((*token), "off", 3))) {
    ret = false;
  }

  (*token) = end;
  return ret;
}

static inline texture_type_t parseTextureType(
    const char **token, texture_type_t default_value = TEXTURE_TYPE_NONE) {
  (*token) += strspn((*token), " \t");
  const char *end = (*token) + strcspn((*token), " \t\r");
  texture_type_t ty = default_value;

  if ((0 == strncmp((*token), "cube_top", strlen("cube_top")))) {
    ty = TEXTURE_TYPE_CUBE_TOP;
  } else if ((0 == strncmp((*token), "cube_bottom", strlen("cube_bottom")))) {
    ty = TEXTURE_TYPE_CUBE_BOTTOM;
  } else if ((0 == strncmp((*token), "cube_left", strlen("cube_left")))) {
    ty = TEXTURE_TYPE_CUBE_LEFT;
  } else if ((0 == strncmp((*token), "cube_right", strlen("cube_right")))) {
    ty = TEXTURE_TYPE_CUBE_RIGHT;
  } else if ((0 == strncmp((*token), "cube_front", strlen("cube_front")))) {
    ty = TEXTURE_TYPE_CUBE_FRONT;
  } else if ((0 == strncmp((*token), "cube_back", strlen("cube_back")))) {
    ty = TEXTURE_TYPE_CUBE_BACK;
  } else if ((0 == strncmp((*token), "sphere", strlen("sphere")))) {
    ty = TEXTURE_TYPE_SPHERE;
  }

  (*token) = end;
  return ty;
}

static tag_sizes parseTagTriple(const char **token) {
  tag_sizes ts;

  (*token) += strspn((*token), " \t");
  ts.num_ints = atoi((*token));
  (*token) += strcspn((*token), "/ \t\r");
  if ((*token)[0] != '/') {
    return ts;
  }

  (*token)++;  // Skip '/'

  (*token) += strspn((*token), " \t");
  ts.num_reals = atoi((*token));
  (*token) += strcspn((*token), "/ \t\r");
  if ((*token)[0] != '/') {
    return ts;
  }
  (*token)++;  // Skip '/'

  ts.num_strings = parseInt(token);

  return ts;
}

// Parse triples with index offsets: i, i/j/k, i//k, i/j
static bool parseTriple(const char **token, int vsize, int vnsize, int vtsize,
                        vertex_index_t *ret, const warning_context &context) {
  if (!ret) {
    return false;
  }

  vertex_index_t vi(-1);

  if (!fixIndex(atoi((*token)), vsize, &vi.v_idx, false, context)) {
    return false;
  }

  (*token) += strcspn((*token), "/ \t\r");
  if ((*token)[0] != '/') {
    (*ret) = vi;
    return true;
  }
  (*token)++;

  // i//k
  if ((*token)[0] == '/') {
    (*token)++;
    if (!fixIndex(atoi((*token)), vnsize, &vi.vn_idx, true, context)) {
      return false;
    }
    (*token) += strcspn((*token), "/ \t\r");
    (*ret) = vi;
    return true;
  }

  // i/j/k or i/j
  if (!fixIndex(atoi((*token)), vtsize, &vi.vt_idx, true, context)) {
    return false;
  }

  (*token) += strcspn((*token), "/ \t\r");
  if ((*token)[0] != '/') {
    (*ret) = vi;
    return true;
  }

  // i/j/k
  (*token)++;  // skip '/'
  if (!fixIndex(atoi((*token)), vnsize, &vi.vn_idx, true, context)) {
    return false;
  }
  (*token) += strcspn((*token), "/ \t\r");

  (*ret) = vi;

  return true;
}

// Parse raw triples: i, i/j/k, i//k, i/j
static vertex_index_t parseRawTriple(const char **token) {
  vertex_index_t vi(static_cast<int>(0));  // 0 is an invalid index in OBJ

  vi.v_idx = atoi((*token));
  (*token) += strcspn((*token), "/ \t\r");
  if ((*token)[0] != '/') {
    return vi;
  }
  (*token)++;

  // i//k
  if ((*token)[0] == '/') {
    (*token)++;
    vi.vn_idx = atoi((*token));
    (*token) += strcspn((*token), "/ \t\r");
    return vi;
  }

  // i/j/k or i/j
  vi.vt_idx = atoi((*token));
  (*token) += strcspn((*token), "/ \t\r");
  if ((*token)[0] != '/') {
    return vi;
  }

  // i/j/k
  (*token)++;  // skip '/'
  vi.vn_idx = atoi((*token));
  (*token) += strcspn((*token), "/ \t\r");
  return vi;
}

// --- Stream-based parse functions ---

static inline std::string sr_parseString(StreamReader &sr) {
  sr.skip_space();
  std::string s;
  while (!sr.eof()) {
    char c = sr.peek();
    if (c == ' ' || c == '\t' || c == '\r' || c == '\n' || c == '\0') break;
    s += c;
    sr.advance(1);
  }
  return s;
}

static inline int sr_parseInt(StreamReader &sr) {
  sr.skip_space();
  const char *start = sr.current_ptr();
  size_t rem = sr.remaining();
  size_t len = 0;
  while (len < rem) {
    char c = start[len];
    if (c == ' ' || c == '\t' || c == '\r' || c == '\n' || c == '\0') break;
    len++;
  }
  int i = 0;
  if (len > 0) {
    char tmp[64];
    size_t copy_len = len < 63 ? len : 63;
    if (copy_len != len) {
      sr.advance(len);
      return 0;
    }
    memcpy(tmp, start, copy_len);
    tmp[copy_len] = '\0';
    errno = 0;
    char *endptr = NULL;
    long val = strtol(tmp, &endptr, 10);
    const bool has_error =
        (errno == ERANGE || endptr == tmp ||
         val > (std::numeric_limits<int>::max)() ||
         val < (std::numeric_limits<int>::min)());
    if (!has_error) {
      i = static_cast<int>(val);
    }
  }
  sr.advance(len);
  return i;
}

static inline real_t sr_parseReal(StreamReader &sr, double default_value = 0.0) {
  sr.skip_space();
  const char *start = sr.current_ptr();
  size_t rem = sr.remaining();
  size_t len = 0;
  while (len < rem) {
    char c = start[len];
    if (c == ' ' || c == '\t' || c == '\r' || c == '\n' || c == '\0') break;
    len++;
  }
  double val = default_value;
  if (len > 0) {
    tryParseDouble(start, start + len, &val);
  }
  sr.advance(len);
  return static_cast<real_t>(val);
}

static inline bool sr_parseReal(StreamReader &sr, real_t *out) {
  sr.skip_space();
  const char *start = sr.current_ptr();
  size_t rem = sr.remaining();
  size_t len = 0;
  while (len < rem) {
    char c = start[len];
    if (c == ' ' || c == '\t' || c == '\r' || c == '\n' || c == '\0') break;
    len++;
  }
  if (len == 0) return false;
  double val;
  bool ret = tryParseDouble(start, start + len, &val);
  if (ret) {
    (*out) = static_cast<real_t>(val);
  }
  sr.advance(len);
  return ret;
}

static inline void sr_parseReal2(real_t *x, real_t *y, StreamReader &sr,
                                 const double default_x = 0.0,
                                 const double default_y = 0.0) {
  (*x) = sr_parseReal(sr, default_x);
  (*y) = sr_parseReal(sr, default_y);
}

static inline void sr_parseReal3(real_t *x, real_t *y, real_t *z,
                                 StreamReader &sr,
                                 const double default_x = 0.0,
                                 const double default_y = 0.0,
                                 const double default_z = 0.0) {
  (*x) = sr_parseReal(sr, default_x);
  (*y) = sr_parseReal(sr, default_y);
  (*z) = sr_parseReal(sr, default_z);
}

static inline int sr_parseVertexWithColor(real_t *x, real_t *y, real_t *z,
                                          real_t *r, real_t *g, real_t *b,
                                          StreamReader &sr,
                                          const double default_x = 0.0,
                                          const double default_y = 0.0,
                                          const double default_z = 0.0) {
  (*x) = sr_parseReal(sr, default_x);
  (*y) = sr_parseReal(sr, default_y);
  (*z) = sr_parseReal(sr, default_z);

  bool has_r = sr_parseReal(sr, r);
  if (!has_r) {
    (*r) = (*g) = (*b) = 1.0;
    return 3;
  }

  bool has_g = sr_parseReal(sr, g);
  if (!has_g) {
    (*g) = (*b) = 1.0;
    return 4;
  }

  bool has_b = sr_parseReal(sr, b);
  if (!has_b) {
    (*r) = (*g) = (*b) = 1.0;
    return 3;
  }

  return 6;
}

// --- Error-reporting overloads ---
// These overloads push clang-style diagnostics into `err` when parsing fails
// and return false so callers can early-return on unrecoverable parse errors.
// The original signatures are preserved above for backward compatibility.

static inline bool sr_parseInt(StreamReader &sr, int *out, std::string *err,
                               const std::string &filename) {
  sr.skip_space();
  const char *start = sr.current_ptr();
  size_t rem = sr.remaining();
  size_t len = 0;
  while (len < rem) {
    char c = start[len];
    if (c == ' ' || c == '\t' || c == '\r' || c == '\n' || c == '\0') break;
    len++;
  }
  if (len == 0) {
    if (err) {
      (*err) += sr.format_error(filename, "expected integer value");
    }
    *out = 0;
    return false;
  }
  char tmp[64];
  size_t copy_len = len < 63 ? len : 63;
  memcpy(tmp, start, copy_len);
  tmp[copy_len] = '\0';
  if (copy_len != len) {
    if (err) {
      (*err) += sr.format_error(filename, "integer value too long");
    }
    *out = 0;
    sr.advance(len);
    return false;
  }
  errno = 0;
  char *endptr = NULL;
  long val = strtol(tmp, &endptr, 10);
  if (errno == ERANGE || val > (std::numeric_limits<int>::max)() ||
      val < (std::numeric_limits<int>::min)()) {
    if (err) {
      (*err) += sr.format_error(filename,
          "integer value out of range, got '" + std::string(tmp) + "'");
    }
    *out = 0;
    sr.advance(len);
    return false;
  }
  if (endptr == tmp || (*endptr != '\0' && *endptr != ' ' && *endptr != '\t')) {
    if (err) {
      (*err) += sr.format_error(filename,
          "expected integer, got '" + std::string(tmp) + "'");
    }
    *out = 0;
    sr.advance(len);
    return false;
  }
  *out = static_cast<int>(val);
  sr.advance(len);
  return true;
}

static inline bool sr_parseReal(StreamReader &sr, real_t *out,
                                 double default_value,
                                 std::string *err,
                                 const std::string &filename) {
  sr.skip_space();
  const char *start = sr.current_ptr();
  size_t rem = sr.remaining();
  size_t len = 0;
  while (len < rem) {
    char c = start[len];
    if (c == ' ' || c == '\t' || c == '\r' || c == '\n' || c == '\0') break;
    len++;
  }
  if (len == 0) {
    // No token to parse — not necessarily an error (e.g. optional component).
    *out = static_cast<real_t>(default_value);
    return true;
  }
  double val;
  if (!tryParseDouble(start, start + len, &val)) {
    if (err) {
      char tmp[64];
      size_t copy_len = len < 63 ? len : 63;
      memcpy(tmp, start, copy_len);
      tmp[copy_len] = '\0';
      (*err) += sr.format_error(filename,
          "expected number, got '" + std::string(tmp) + "'");
    }
    *out = static_cast<real_t>(default_value);
    sr.advance(len);
    return false;
  }
  *out = static_cast<real_t>(val);
  sr.advance(len);
  return true;
}

static inline bool sr_parseReal2(real_t *x, real_t *y, StreamReader &sr,
                                  std::string *err,
                                  const std::string &filename,
                                  const double default_x = 0.0,
                                  const double default_y = 0.0) {
  if (!sr_parseReal(sr, x, default_x, err, filename)) return false;
  if (!sr_parseReal(sr, y, default_y, err, filename)) return false;
  return true;
}

static inline bool sr_parseReal3(real_t *x, real_t *y, real_t *z,
                                  StreamReader &sr,
                                  std::string *err,
                                  const std::string &filename,
                                  const double default_x = 0.0,
                                  const double default_y = 0.0,
                                  const double default_z = 0.0) {
  if (!sr_parseReal(sr, x, default_x, err, filename)) return false;
  if (!sr_parseReal(sr, y, default_y, err, filename)) return false;
  if (!sr_parseReal(sr, z, default_z, err, filename)) return false;
  return true;
}

// Returns number of components parsed (3, 4, or 6) on success, -1 on error.
static inline int sr_parseVertexWithColor(real_t *x, real_t *y, real_t *z,
                                          real_t *r, real_t *g, real_t *b,
                                          StreamReader &sr,
                                          std::string *err,
                                          const std::string &filename,
                                          const double default_x = 0.0,
                                          const double default_y = 0.0,
                                          const double default_z = 0.0) {
  if (!sr_parseReal(sr, x, default_x, err, filename)) return -1;
  if (!sr_parseReal(sr, y, default_y, err, filename)) return -1;
  if (!sr_parseReal(sr, z, default_z, err, filename)) return -1;

  bool has_r = sr_parseReal(sr, r);
  if (!has_r) {
    (*r) = (*g) = (*b) = 1.0;
    return 3;
  }

  bool has_g = sr_parseReal(sr, g);
  if (!has_g) {
    (*g) = (*b) = 1.0;
    return 4;
  }

  bool has_b = sr_parseReal(sr, b);
  if (!has_b) {
    (*r) = (*g) = (*b) = 1.0;
    return 3;
  }

  return 6;
}

static inline int sr_parseIntNoSkip(StreamReader &sr);

// Advance past remaining characters in a tag triple field (stops at '/', whitespace, or line end).
static inline void sr_skipTagField(StreamReader &sr) {
  while (!sr.eof() && !sr.at_line_end() && !IS_SPACE(sr.peek()) &&
         sr.peek() != '/') {
    sr.advance(1);
  }
}

static tag_sizes sr_parseTagTriple(StreamReader &sr) {
  tag_sizes ts;

  sr.skip_space();
  ts.num_ints = sr_parseIntNoSkip(sr);
  sr_skipTagField(sr);
  if (!sr.eof() && sr.peek() == '/') {
    sr.advance(1);
    sr.skip_space();
    ts.num_reals = sr_parseIntNoSkip(sr);
    sr_skipTagField(sr);
    if (!sr.eof() && sr.peek() == '/') {
      sr.advance(1);
      ts.num_strings = sr_parseInt(sr);
    }
  }
  return ts;
}

static inline int sr_parseIntNoSkip(StreamReader &sr) {
  const char *start = sr.current_ptr();
  size_t rem = sr.remaining();
  size_t len = 0;
  if (len < rem && (start[len] == '+' || start[len] == '-')) len++;
  while (len < rem && start[len] >= '0' && start[len] <= '9') len++;
  int i = 0;
  if (len > 0) {
    char tmp[64];
    size_t copy_len = len < 63 ? len : 63;
    if (copy_len != len) {
      sr.advance(len);
      return 0;
    }
    memcpy(tmp, start, copy_len);
    tmp[copy_len] = '\0';
    errno = 0;
    char *endptr = NULL;
    long val = strtol(tmp, &endptr, 10);
    if (errno == 0 && endptr != tmp && *endptr == '\0' &&
        val <= (std::numeric_limits<int>::max)() &&
        val >= (std::numeric_limits<int>::min)()) {
      i = static_cast<int>(val);
    }
  }
  sr.advance(len);
  return i;
}

static inline void sr_skipUntil(StreamReader &sr, const char *delims) {
  while (!sr.eof()) {
    char c = sr.peek();
    for (const char *d = delims; *d; d++) {
      if (c == *d) return;
    }
    sr.advance(1);
  }
}

static bool sr_parseTriple(StreamReader &sr, int vsize, int vnsize, int vtsize,
                           vertex_index_t *ret, const warning_context &context) {
  if (!ret) return false;

  vertex_index_t vi(-1);

  sr.skip_space();
  if (!fixIndex(sr_parseIntNoSkip(sr), vsize, &vi.v_idx, false, context)) {
    return false;
  }

  sr_skipUntil(sr, "/ \t\r\n");
  if (sr.eof() || sr.peek() != '/') {
    (*ret) = vi;
    return true;
  }
  sr.advance(1);

  // i//k
  if (!sr.eof() && sr.peek() == '/') {
    sr.advance(1);
    if (!fixIndex(sr_parseIntNoSkip(sr), vnsize, &vi.vn_idx, true, context)) {
      return false;
    }
    sr_skipUntil(sr, "/ \t\r\n");
    (*ret) = vi;
    return true;
  }

  // i/j/k or i/j
  if (!fixIndex(sr_parseIntNoSkip(sr), vtsize, &vi.vt_idx, true, context)) {
    return false;
  }

  sr_skipUntil(sr, "/ \t\r\n");
  if (sr.eof() || sr.peek() != '/') {
    (*ret) = vi;
    return true;
  }

  // i/j/k
  sr.advance(1);
  if (!fixIndex(sr_parseIntNoSkip(sr), vnsize, &vi.vn_idx, true, context)) {
    return false;
  }
  sr_skipUntil(sr, "/ \t\r\n");

  (*ret) = vi;
  return true;
}

static vertex_index_t sr_parseRawTriple(StreamReader &sr) {
  vertex_index_t vi(static_cast<int>(0));

  sr.skip_space();
  vi.v_idx = sr_parseIntNoSkip(sr);
  sr_skipUntil(sr, "/ \t\r\n");
  if (sr.eof() || sr.peek() != '/') return vi;
  sr.advance(1);

  // i//k
  if (!sr.eof() && sr.peek() == '/') {
    sr.advance(1);
    vi.vn_idx = sr_parseIntNoSkip(sr);
    sr_skipUntil(sr, "/ \t\r\n");
    return vi;
  }

  // i/j/k or i/j
  vi.vt_idx = sr_parseIntNoSkip(sr);
  sr_skipUntil(sr, "/ \t\r\n");
  if (sr.eof() || sr.peek() != '/') return vi;

  sr.advance(1);
  vi.vn_idx = sr_parseIntNoSkip(sr);
  sr_skipUntil(sr, "/ \t\r\n");
  return vi;
}

bool ParseTextureNameAndOption(std::string *texname, texture_option_t *texopt,
                               const char *linebuf) {
  // @todo { write more robust lexer and parser. }
  bool found_texname = false;
  std::string texture_name;

  const char *token = linebuf;  // Assume line ends with NULL

  while (!IS_NEW_LINE((*token))) {
    token += strspn(token, " \t");  // skip space
    if ((0 == strncmp(token, "-blendu", 7)) && IS_SPACE((token[7]))) {
      token += 8;
      texopt->blendu = parseOnOff(&token, /* default */ true);
    } else if ((0 == strncmp(token, "-blendv", 7)) && IS_SPACE((token[7]))) {
      token += 8;
      texopt->blendv = parseOnOff(&token, /* default */ true);
    } else if ((0 == strncmp(token, "-clamp", 6)) && IS_SPACE((token[6]))) {
      token += 7;
      texopt->clamp = parseOnOff(&token, /* default */ true);
    } else if ((0 == strncmp(token, "-boost", 6)) && IS_SPACE((token[6]))) {
      token += 7;
      texopt->sharpness = parseReal(&token, 1.0);
    } else if ((0 == strncmp(token, "-bm", 3)) && IS_SPACE((token[3]))) {
      token += 4;
      texopt->bump_multiplier = parseReal(&token, 1.0);
    } else if ((0 == strncmp(token, "-o", 2)) && IS_SPACE((token[2]))) {
      token += 3;
      parseReal3(&(texopt->origin_offset[0]), &(texopt->origin_offset[1]),
                 &(texopt->origin_offset[2]), &token);
    } else if ((0 == strncmp(token, "-s", 2)) && IS_SPACE((token[2]))) {
      token += 3;
      parseReal3(&(texopt->scale[0]), &(texopt->scale[1]), &(texopt->scale[2]),
                 &token, 1.0, 1.0, 1.0);
    } else if ((0 == strncmp(token, "-t", 2)) && IS_SPACE((token[2]))) {
      token += 3;
      parseReal3(&(texopt->turbulence[0]), &(texopt->turbulence[1]),
                 &(texopt->turbulence[2]), &token);
    } else if ((0 == strncmp(token, "-type", 5)) && IS_SPACE((token[5]))) {
      token += 5;
      texopt->type = parseTextureType((&token), TEXTURE_TYPE_NONE);
    } else if ((0 == strncmp(token, "-texres", 7)) && IS_SPACE((token[7]))) {
      token += 7;
      // TODO(syoyo): Check if arg is int type.
      texopt->texture_resolution = parseInt(&token);
    } else if ((0 == strncmp(token, "-imfchan", 8)) && IS_SPACE((token[8]))) {
      token += 9;
      token += strspn(token, " \t");
      const char *end = token + strcspn(token, " \t\r");
      if ((end - token) == 1) {  // Assume one char for -imfchan
        texopt->imfchan = (*token);
      }
      token = end;
    } else if ((0 == strncmp(token, "-mm", 3)) && IS_SPACE((token[3]))) {
      token += 4;
      parseReal2(&(texopt->brightness), &(texopt->contrast), &token, 0.0, 1.0);
    } else if ((0 == strncmp(token, "-colorspace", 11)) &&
               IS_SPACE((token[11]))) {
      token += 12;
      texopt->colorspace = parseString(&token);
    } else {
// Assume texture filename
#if 0
      size_t len = strcspn(token, " \t\r");  // untile next space
      texture_name = std::string(token, token + len);
      token += len;

      token += strspn(token, " \t");  // skip space
#else
      // Read filename until line end to parse filename containing whitespace
      // TODO(syoyo): Support parsing texture option flag after the filename.
      texture_name = std::string(token);
      token += texture_name.length();
#endif

      found_texname = true;
    }
  }

  if (found_texname) {
    (*texname) = texture_name;
    return true;
  } else {
    return false;
  }
}

static void InitTexOpt(texture_option_t *texopt, const bool is_bump) {
  if (is_bump) {
    texopt->imfchan = 'l';
  } else {
    texopt->imfchan = 'm';
  }
  texopt->bump_multiplier = static_cast<real_t>(1.0);
  texopt->clamp = false;
  texopt->blendu = true;
  texopt->blendv = true;
  texopt->sharpness = static_cast<real_t>(1.0);
  texopt->brightness = static_cast<real_t>(0.0);
  texopt->contrast = static_cast<real_t>(1.0);
  texopt->origin_offset[0] = static_cast<real_t>(0.0);
  texopt->origin_offset[1] = static_cast<real_t>(0.0);
  texopt->origin_offset[2] = static_cast<real_t>(0.0);
  texopt->scale[0] = static_cast<real_t>(1.0);
  texopt->scale[1] = static_cast<real_t>(1.0);
  texopt->scale[2] = static_cast<real_t>(1.0);
  texopt->turbulence[0] = static_cast<real_t>(0.0);
  texopt->turbulence[1] = static_cast<real_t>(0.0);
  texopt->turbulence[2] = static_cast<real_t>(0.0);
  texopt->texture_resolution = -1;
  texopt->type = TEXTURE_TYPE_NONE;
}

static void InitMaterial(material_t *material) {
  InitTexOpt(&material->ambient_texopt, /* is_bump */ false);
  InitTexOpt(&material->diffuse_texopt, /* is_bump */ false);
  InitTexOpt(&material->specular_texopt, /* is_bump */ false);
  InitTexOpt(&material->specular_highlight_texopt, /* is_bump */ false);
  InitTexOpt(&material->bump_texopt, /* is_bump */ true);
  InitTexOpt(&material->displacement_texopt, /* is_bump */ false);
  InitTexOpt(&material->alpha_texopt, /* is_bump */ false);
  InitTexOpt(&material->reflection_texopt, /* is_bump */ false);
  InitTexOpt(&material->roughness_texopt, /* is_bump */ false);
  InitTexOpt(&material->metallic_texopt, /* is_bump */ false);
  InitTexOpt(&material->sheen_texopt, /* is_bump */ false);
  InitTexOpt(&material->emissive_texopt, /* is_bump */ false);
  InitTexOpt(&material->normal_texopt,
             /* is_bump */ false);  // @fixme { is_bump will be true? }
  material->name = "";
  material->ambient_texname = "";
  material->diffuse_texname = "";
  material->specular_texname = "";
  material->specular_highlight_texname = "";
  material->bump_texname = "";
  material->displacement_texname = "";
  material->reflection_texname = "";
  material->alpha_texname = "";
  for (int i = 0; i < 3; i++) {
    material->ambient[i] = static_cast<real_t>(0.0);
    material->diffuse[i] = static_cast<real_t>(0.0);
    material->specular[i] = static_cast<real_t>(0.0);
    material->transmittance[i] = static_cast<real_t>(0.0);
    material->emission[i] = static_cast<real_t>(0.0);
  }
  material->illum = 0;
  material->dissolve = static_cast<real_t>(1.0);
  material->shininess = static_cast<real_t>(1.0);
  material->ior = static_cast<real_t>(1.0);

  material->roughness = static_cast<real_t>(0.0);
  material->metallic = static_cast<real_t>(0.0);
  material->sheen = static_cast<real_t>(0.0);
  material->clearcoat_thickness = static_cast<real_t>(0.0);
  material->clearcoat_roughness = static_cast<real_t>(0.0);
  material->anisotropy_rotation = static_cast<real_t>(0.0);
  material->anisotropy = static_cast<real_t>(0.0);
  material->roughness_texname = "";
  material->metallic_texname = "";
  material->sheen_texname = "";
  material->emissive_texname = "";
  material->normal_texname = "";

  material->unknown_parameter.clear();
}

// code from https://wrf.ecse.rpi.edu//Research/Short_Notes/pnpoly.html
template <typename T>
static int pnpoly(int nvert, T *vertx, T *verty, T testx, T testy) {
  int i, j, c = 0;
  for (i = 0, j = nvert - 1; i < nvert; j = i++) {
    if (((verty[i] > testy) != (verty[j] > testy)) &&
        (testx <
         (vertx[j] - vertx[i]) * (testy - verty[i]) / (verty[j] - verty[i]) +
             vertx[i]))
      c = !c;
  }
  return c;
}

struct TinyObjPoint {
  real_t x, y, z;
  TinyObjPoint() : x(0), y(0), z(0) {}
  TinyObjPoint(real_t x_, real_t y_, real_t z_) : x(x_), y(y_), z(z_) {}
};

inline TinyObjPoint cross(const TinyObjPoint &v1, const TinyObjPoint &v2) {
  return TinyObjPoint(v1.y * v2.z - v1.z * v2.y, v1.z * v2.x - v1.x * v2.z,
                      v1.x * v2.y - v1.y * v2.x);
}

inline real_t dot(const TinyObjPoint &v1, const TinyObjPoint &v2) {
  return (v1.x * v2.x + v1.y * v2.y + v1.z * v2.z);
}

inline real_t GetLength(TinyObjPoint &e) {
  return std::sqrt(e.x * e.x + e.y * e.y + e.z * e.z);
}

inline TinyObjPoint Normalize(TinyObjPoint e) {
  real_t len = GetLength(e);
  if (len <= real_t(0)) return TinyObjPoint(real_t(0), real_t(0), real_t(0));
  real_t inv_length = real_t(1) / len;
  return TinyObjPoint(e.x * inv_length, e.y * inv_length, e.z * inv_length);
}

inline TinyObjPoint WorldToLocal(const TinyObjPoint &a, const TinyObjPoint &u,
                                 const TinyObjPoint &v, const TinyObjPoint &w) {
  return TinyObjPoint(dot(a, u), dot(a, v), dot(a, w));
}

// TODO(syoyo): refactor function.
static bool exportGroupsToShape(shape_t *shape, const PrimGroup &prim_group,
                                const std::vector<tag_t> &tags,
                                const int material_id, const std::string &name,
                                bool triangulate, const std::vector<real_t> &v,
                                std::string *warn) {
  if (prim_group.IsEmpty()) {
    return false;
  }

  shape->name = name;

  // polygon
  if (!prim_group.faceGroup.empty()) {
    // Flatten vertices and indices
    for (size_t i = 0; i < prim_group.faceGroup.size(); i++) {
      const face_t &face = prim_group.faceGroup[i];

      size_t npolys = face.vertex_indices.size();

      if (npolys < 3) {
        // Face must have 3+ vertices.
        if (warn) {
          (*warn) += "Degenerated face found\n.";
        }
        continue;
      }

      if (triangulate && npolys != 3) {
        if (npolys == 4) {
          vertex_index_t i0 = face.vertex_indices[0];
          vertex_index_t i1 = face.vertex_indices[1];
          vertex_index_t i2 = face.vertex_indices[2];
          vertex_index_t i3 = face.vertex_indices[3];

          if (i0.v_idx < 0 || i1.v_idx < 0 || i2.v_idx < 0 || i3.v_idx < 0) {
            if (warn) {
              (*warn) += "Face with invalid vertex index found.\n";
            }
            continue;
          }

          size_t vi0 = size_t(i0.v_idx);
          size_t vi1 = size_t(i1.v_idx);
          size_t vi2 = size_t(i2.v_idx);
          size_t vi3 = size_t(i3.v_idx);

          if (((3 * vi0 + 2) >= v.size()) || ((3 * vi1 + 2) >= v.size()) ||
              ((3 * vi2 + 2) >= v.size()) || ((3 * vi3 + 2) >= v.size())) {
            // Invalid triangle.
            // FIXME(syoyo): Is it ok to simply skip this invalid triangle?
            if (warn) {
              (*warn) += "Face with invalid vertex index found.\n";
            }
            continue;
          }

          real_t v0x = v[vi0 * 3 + 0];
          real_t v0y = v[vi0 * 3 + 1];
          real_t v0z = v[vi0 * 3 + 2];
          real_t v1x = v[vi1 * 3 + 0];
          real_t v1y = v[vi1 * 3 + 1];
          real_t v1z = v[vi1 * 3 + 2];
          real_t v2x = v[vi2 * 3 + 0];
          real_t v2y = v[vi2 * 3 + 1];
          real_t v2z = v[vi2 * 3 + 2];
          real_t v3x = v[vi3 * 3 + 0];
          real_t v3y = v[vi3 * 3 + 1];
          real_t v3z = v[vi3 * 3 + 2];

          // There are two candidates to split the quad into two triangles.
          //
          // Choose the shortest edge.
          // TODO: Is it better to determine the edge to split by calculating
          // the area of each triangle?
          //
          // +---+
          // |\  |
          // | \ |
          // |  \|
          // +---+
          //
          // +---+
          // |  /|
          // | / |
          // |/  |
          // +---+

          real_t e02x = v2x - v0x;
          real_t e02y = v2y - v0y;
          real_t e02z = v2z - v0z;
          real_t e13x = v3x - v1x;
          real_t e13y = v3y - v1y;
          real_t e13z = v3z - v1z;

          real_t sqr02 = e02x * e02x + e02y * e02y + e02z * e02z;
          real_t sqr13 = e13x * e13x + e13y * e13y + e13z * e13z;

          index_t idx0, idx1, idx2, idx3;

          idx0.vertex_index = i0.v_idx;
          idx0.normal_index = i0.vn_idx;
          idx0.texcoord_index = i0.vt_idx;
          idx1.vertex_index = i1.v_idx;
          idx1.normal_index = i1.vn_idx;
          idx1.texcoord_index = i1.vt_idx;
          idx2.vertex_index = i2.v_idx;
          idx2.normal_index = i2.vn_idx;
          idx2.texcoord_index = i2.vt_idx;
          idx3.vertex_index = i3.v_idx;
          idx3.normal_index = i3.vn_idx;
          idx3.texcoord_index = i3.vt_idx;

          if (sqr02 < sqr13) {
            // [0, 1, 2], [0, 2, 3]
            shape->mesh.indices.push_back(idx0);
            shape->mesh.indices.push_back(idx1);
            shape->mesh.indices.push_back(idx2);

            shape->mesh.indices.push_back(idx0);
            shape->mesh.indices.push_back(idx2);
            shape->mesh.indices.push_back(idx3);
          } else {
            // [0, 1, 3], [1, 2, 3]
            shape->mesh.indices.push_back(idx0);
            shape->mesh.indices.push_back(idx1);
            shape->mesh.indices.push_back(idx3);

            shape->mesh.indices.push_back(idx1);
            shape->mesh.indices.push_back(idx2);
            shape->mesh.indices.push_back(idx3);
          }

          // Two triangle faces
          shape->mesh.num_face_vertices.push_back(3);
          shape->mesh.num_face_vertices.push_back(3);

          shape->mesh.material_ids.push_back(material_id);
          shape->mesh.material_ids.push_back(material_id);

          shape->mesh.smoothing_group_ids.push_back(face.smoothing_group_id);
          shape->mesh.smoothing_group_ids.push_back(face.smoothing_group_id);

        } else {
#ifdef TINYOBJLOADER_USE_MAPBOX_EARCUT
          // Validate all vertex indices before accessing the vertex array.
          {
            bool valid_poly = true;
            for (size_t k = 0; k < npolys; ++k) {
              size_t vi = size_t(face.vertex_indices[k].v_idx);
              if ((3 * vi + 2) >= v.size()) {
                valid_poly = false;
                break;
              }
            }
            if (!valid_poly) {
              if (warn) {
                (*warn) += "Face with invalid vertex index found.\n";
              }
              continue;
            }
          }

          vertex_index_t i0 = face.vertex_indices[0];
          vertex_index_t i0_2 = i0;

          // TMW change: Find the normal axis of the polygon using Newell's
          // method
          TinyObjPoint n;
          for (size_t k = 0; k < npolys; ++k) {
            i0 = face.vertex_indices[k % npolys];
            size_t vi0 = size_t(i0.v_idx);

            size_t j = (k + 1) % npolys;
            i0_2 = face.vertex_indices[j];
            size_t vi0_2 = size_t(i0_2.v_idx);

            real_t v0x = v[vi0 * 3 + 0];
            real_t v0y = v[vi0 * 3 + 1];
            real_t v0z = v[vi0 * 3 + 2];

            real_t v0x_2 = v[vi0_2 * 3 + 0];
            real_t v0y_2 = v[vi0_2 * 3 + 1];
            real_t v0z_2 = v[vi0_2 * 3 + 2];

            const TinyObjPoint point1(v0x, v0y, v0z);
            const TinyObjPoint point2(v0x_2, v0y_2, v0z_2);

            TinyObjPoint a(point1.x - point2.x, point1.y - point2.y,
                           point1.z - point2.z);
            TinyObjPoint b(point1.x + point2.x, point1.y + point2.y,
                           point1.z + point2.z);

            n.x += (a.y * b.z);
            n.y += (a.z * b.x);
            n.z += (a.x * b.y);
          }
          real_t length_n = GetLength(n);
          // Check if zero length normal
          if (length_n <= 0) {
            continue;
          }
          // Negative is to flip the normal to the correct direction
          real_t inv_length = -real_t(1.0) / length_n;
          n.x *= inv_length;
          n.y *= inv_length;
          n.z *= inv_length;

          TinyObjPoint axis_w, axis_v, axis_u;
          axis_w = n;
          TinyObjPoint a;
          if (std::fabs(axis_w.x) > real_t(0.9999999)) {
            a = TinyObjPoint(0, 1, 0);
          } else {
            a = TinyObjPoint(1, 0, 0);
          }
          axis_v = Normalize(cross(axis_w, a));
          axis_u = cross(axis_w, axis_v);
          using Point = std::array<real_t, 2>;

          // first polyline define the main polygon.
          // following polylines define holes(not used in tinyobj).
          std::vector<std::vector<Point> > polygon;

          std::vector<Point> polyline;

          // TMW change: Find best normal and project v0x and v0y to those
          // coordinates, instead of picking a plane aligned with an axis (which
          // can flip polygons).

          // Fill polygon data(facevarying vertices).
          for (size_t k = 0; k < npolys; k++) {
            i0 = face.vertex_indices[k];
            size_t vi0 = size_t(i0.v_idx);

            assert(((3 * vi0 + 2) < v.size()));

            real_t v0x = v[vi0 * 3 + 0];
            real_t v0y = v[vi0 * 3 + 1];
            real_t v0z = v[vi0 * 3 + 2];

            TinyObjPoint polypoint(v0x, v0y, v0z);
            TinyObjPoint loc = WorldToLocal(polypoint, axis_u, axis_v, axis_w);

            polyline.push_back({loc.x, loc.y});
          }

          polygon.push_back(polyline);
          std::vector<uint32_t> indices = mapbox::earcut<uint32_t>(polygon);
          // => result = 3 * faces, clockwise

          assert(indices.size() % 3 == 0);

          // Reconstruct vertex_index_t
          for (size_t k = 0; k < indices.size() / 3; k++) {
            {
              index_t idx0, idx1, idx2;
              idx0.vertex_index = face.vertex_indices[indices[3 * k + 0]].v_idx;
              idx0.normal_index =
                  face.vertex_indices[indices[3 * k + 0]].vn_idx;
              idx0.texcoord_index =
                  face.vertex_indices[indices[3 * k + 0]].vt_idx;
              idx1.vertex_index = face.vertex_indices[indices[3 * k + 1]].v_idx;
              idx1.normal_index =
                  face.vertex_indices[indices[3 * k + 1]].vn_idx;
              idx1.texcoord_index =
                  face.vertex_indices[indices[3 * k + 1]].vt_idx;
              idx2.vertex_index = face.vertex_indices[indices[3 * k + 2]].v_idx;
              idx2.normal_index =
                  face.vertex_indices[indices[3 * k + 2]].vn_idx;
              idx2.texcoord_index =
                  face.vertex_indices[indices[3 * k + 2]].vt_idx;

              shape->mesh.indices.push_back(idx0);
              shape->mesh.indices.push_back(idx1);
              shape->mesh.indices.push_back(idx2);

              shape->mesh.num_face_vertices.push_back(3);
              shape->mesh.material_ids.push_back(material_id);
              shape->mesh.smoothing_group_ids.push_back(
                  face.smoothing_group_id);
            }
          }

#else  // Built-in ear clipping triangulation
          vertex_index_t i0 = face.vertex_indices[0];
          vertex_index_t i1(-1);
          vertex_index_t i2 = face.vertex_indices[1];

          // find the two axes to work in
          size_t axes[2] = {1, 2};
          for (size_t k = 0; k < npolys; ++k) {
            i0 = face.vertex_indices[(k + 0) % npolys];
            i1 = face.vertex_indices[(k + 1) % npolys];
            i2 = face.vertex_indices[(k + 2) % npolys];
            size_t vi0 = size_t(i0.v_idx);
            size_t vi1 = size_t(i1.v_idx);
            size_t vi2 = size_t(i2.v_idx);

            if (((3 * vi0 + 2) >= v.size()) || ((3 * vi1 + 2) >= v.size()) ||
                ((3 * vi2 + 2) >= v.size())) {
              // Invalid triangle.
              // FIXME(syoyo): Is it ok to simply skip this invalid triangle?
              continue;
            }
            real_t v0x = v[vi0 * 3 + 0];
            real_t v0y = v[vi0 * 3 + 1];
            real_t v0z = v[vi0 * 3 + 2];
            real_t v1x = v[vi1 * 3 + 0];
            real_t v1y = v[vi1 * 3 + 1];
            real_t v1z = v[vi1 * 3 + 2];
            real_t v2x = v[vi2 * 3 + 0];
            real_t v2y = v[vi2 * 3 + 1];
            real_t v2z = v[vi2 * 3 + 2];
            real_t e0x = v1x - v0x;
            real_t e0y = v1y - v0y;
            real_t e0z = v1z - v0z;
            real_t e1x = v2x - v1x;
            real_t e1y = v2y - v1y;
            real_t e1z = v2z - v1z;
            real_t cx = std::fabs(e0y * e1z - e0z * e1y);
            real_t cy = std::fabs(e0z * e1x - e0x * e1z);
            real_t cz = std::fabs(e0x * e1y - e0y * e1x);
            const real_t epsilon = std::numeric_limits<real_t>::epsilon();
            // std::cout << "cx " << cx << ", cy " << cy << ", cz " << cz <<
            // "\n";
            if (cx > epsilon || cy > epsilon || cz > epsilon) {
              // std::cout << "corner\n";
              // found a corner
              if (cx > cy && cx > cz) {
                // std::cout << "pattern0\n";
              } else {
                // std::cout << "axes[0] = 0\n";
                axes[0] = 0;
                if (cz > cx && cz > cy) {
                  // std::cout << "axes[1] = 1\n";
                  axes[1] = 1;
                }
              }
              break;
            }
          }

          face_t remainingFace = face;  // copy
          size_t guess_vert = 0;
          vertex_index_t ind[3];
          real_t vx[3];
          real_t vy[3];

          // How many iterations can we do without decreasing the remaining
          // vertices.
          size_t remainingIterations = face.vertex_indices.size();
          size_t previousRemainingVertices =
              remainingFace.vertex_indices.size();

          while (remainingFace.vertex_indices.size() > 3 &&
                 remainingIterations > 0) {
            // std::cout << "remainingIterations " << remainingIterations <<
            // "\n";

            npolys = remainingFace.vertex_indices.size();
            if (guess_vert >= npolys) {
              guess_vert -= npolys;
            }

            if (previousRemainingVertices != npolys) {
              // The number of remaining vertices decreased. Reset counters.
              previousRemainingVertices = npolys;
              remainingIterations = npolys;
            } else {
              // We didn't consume a vertex on previous iteration, reduce the
              // available iterations.
              remainingIterations--;
            }

            for (size_t k = 0; k < 3; k++) {
              ind[k] = remainingFace.vertex_indices[(guess_vert + k) % npolys];
              size_t vi = size_t(ind[k].v_idx);
              if (((vi * 3 + axes[0]) >= v.size()) ||
                  ((vi * 3 + axes[1]) >= v.size())) {
                // ???
                vx[k] = static_cast<real_t>(0.0);
                vy[k] = static_cast<real_t>(0.0);
              } else {
                vx[k] = v[vi * 3 + axes[0]];
                vy[k] = v[vi * 3 + axes[1]];
              }
            }

            //
            // area is calculated per face
            //
            real_t e0x = vx[1] - vx[0];
            real_t e0y = vy[1] - vy[0];
            real_t e1x = vx[2] - vx[1];
            real_t e1y = vy[2] - vy[1];
            real_t cross = e0x * e1y - e0y * e1x;
            // std::cout << "axes = " << axes[0] << ", " << axes[1] << "\n";
            // std::cout << "e0x, e0y, e1x, e1y " << e0x << ", " << e0y << ", "
            // << e1x << ", " << e1y << "\n";

            real_t area =
                (vx[0] * vy[1] - vy[0] * vx[1]) * static_cast<real_t>(0.5);
            // std::cout << "cross " << cross << ", area " << area << "\n";
            // if an internal angle
            if (cross * area < static_cast<real_t>(0.0)) {
              // std::cout << "internal \n";
              guess_vert += 1;
              // std::cout << "guess vert : " << guess_vert << "\n";
              continue;
            }

            // check all other verts in case they are inside this triangle
            bool overlap = false;
            for (size_t otherVert = 3; otherVert < npolys; ++otherVert) {
              size_t idx = (guess_vert + otherVert) % npolys;

              if (idx >= remainingFace.vertex_indices.size()) {
                // std::cout << "???0\n";
                // ???
                continue;
              }

              size_t ovi = size_t(remainingFace.vertex_indices[idx].v_idx);

              if (((ovi * 3 + axes[0]) >= v.size()) ||
                  ((ovi * 3 + axes[1]) >= v.size())) {
                // std::cout << "???1\n";
                // ???
                continue;
              }
              real_t tx = v[ovi * 3 + axes[0]];
              real_t ty = v[ovi * 3 + axes[1]];
              if (pnpoly(3, vx, vy, tx, ty)) {
                // std::cout << "overlap\n";
                overlap = true;
                break;
              }
            }

            if (overlap) {
              // std::cout << "overlap2\n";
              guess_vert += 1;
              continue;
            }

            // this triangle is an ear
            {
              index_t idx0, idx1, idx2;
              idx0.vertex_index = ind[0].v_idx;
              idx0.normal_index = ind[0].vn_idx;
              idx0.texcoord_index = ind[0].vt_idx;
              idx1.vertex_index = ind[1].v_idx;
              idx1.normal_index = ind[1].vn_idx;
              idx1.texcoord_index = ind[1].vt_idx;
              idx2.vertex_index = ind[2].v_idx;
              idx2.normal_index = ind[2].vn_idx;
              idx2.texcoord_index = ind[2].vt_idx;

              shape->mesh.indices.push_back(idx0);
              shape->mesh.indices.push_back(idx1);
              shape->mesh.indices.push_back(idx2);

              shape->mesh.num_face_vertices.push_back(3);
              shape->mesh.material_ids.push_back(material_id);
              shape->mesh.smoothing_group_ids.push_back(
                  face.smoothing_group_id);
            }

            // remove v1 from the list
            size_t removed_vert_index = (guess_vert + 1) % npolys;
            while (removed_vert_index + 1 < npolys) {
              remainingFace.vertex_indices[removed_vert_index] =
                  remainingFace.vertex_indices[removed_vert_index + 1];
              removed_vert_index += 1;
            }
            remainingFace.vertex_indices.pop_back();
          }

          // std::cout << "remainingFace.vi.size = " <<
          // remainingFace.vertex_indices.size() << "\n";
          if (remainingFace.vertex_indices.size() == 3) {
            i0 = remainingFace.vertex_indices[0];
            i1 = remainingFace.vertex_indices[1];
            i2 = remainingFace.vertex_indices[2];
            {
              index_t idx0, idx1, idx2;
              idx0.vertex_index = i0.v_idx;
              idx0.normal_index = i0.vn_idx;
              idx0.texcoord_index = i0.vt_idx;
              idx1.vertex_index = i1.v_idx;
              idx1.normal_index = i1.vn_idx;
              idx1.texcoord_index = i1.vt_idx;
              idx2.vertex_index = i2.v_idx;
              idx2.normal_index = i2.vn_idx;
              idx2.texcoord_index = i2.vt_idx;

              shape->mesh.indices.push_back(idx0);
              shape->mesh.indices.push_back(idx1);
              shape->mesh.indices.push_back(idx2);

              shape->mesh.num_face_vertices.push_back(3);
              shape->mesh.material_ids.push_back(material_id);
              shape->mesh.smoothing_group_ids.push_back(
                  face.smoothing_group_id);
            }
          }
#endif
        }  // npolys
      } else {
        for (size_t k = 0; k < npolys; k++) {
          index_t idx;
          idx.vertex_index = face.vertex_indices[k].v_idx;
          idx.normal_index = face.vertex_indices[k].vn_idx;
          idx.texcoord_index = face.vertex_indices[k].vt_idx;
          shape->mesh.indices.push_back(idx);
        }

        shape->mesh.num_face_vertices.push_back(
            static_cast<unsigned int>(npolys));
        shape->mesh.material_ids.push_back(material_id);  // per face
        shape->mesh.smoothing_group_ids.push_back(
            face.smoothing_group_id);  // per face
      }
    }

    shape->mesh.tags = tags;
  }

  // line
  if (!prim_group.lineGroup.empty()) {
    // Flatten indices
    for (size_t i = 0; i < prim_group.lineGroup.size(); i++) {
      for (size_t j = 0; j < prim_group.lineGroup[i].vertex_indices.size();
           j++) {
        const vertex_index_t &vi = prim_group.lineGroup[i].vertex_indices[j];

        index_t idx;
        idx.vertex_index = vi.v_idx;
        idx.normal_index = vi.vn_idx;
        idx.texcoord_index = vi.vt_idx;

        shape->lines.indices.push_back(idx);
      }

      shape->lines.num_line_vertices.push_back(
          int(prim_group.lineGroup[i].vertex_indices.size()));
    }
  }

  // points
  if (!prim_group.pointsGroup.empty()) {
    // Flatten & convert indices
    for (size_t i = 0; i < prim_group.pointsGroup.size(); i++) {
      for (size_t j = 0; j < prim_group.pointsGroup[i].vertex_indices.size();
           j++) {
        const vertex_index_t &vi = prim_group.pointsGroup[i].vertex_indices[j];

        index_t idx;
        idx.vertex_index = vi.v_idx;
        idx.normal_index = vi.vn_idx;
        idx.texcoord_index = vi.vt_idx;

        shape->points.indices.push_back(idx);
      }
    }
  }

  return true;
}

// Split a string with specified delimiter character and escape character.
// https://rosettacode.org/wiki/Tokenize_a_string_with_escaping#C.2B.2B
static void SplitString(const std::string &s, char delim, char escape,
                        std::vector<std::string> &elems) {
  std::string token;

  bool escaping = false;
  for (size_t i = 0; i < s.size(); ++i) {
    char ch = s[i];
    if (escaping) {
      escaping = false;
    } else if (ch == escape) {
      if ((i + 1) < s.size()) {
        const char next = s[i + 1];
        if ((next == delim) || (next == escape)) {
          escaping = true;
          continue;
        }
      }
    } else if (ch == delim) {
      if (!token.empty()) {
        elems.push_back(token);
      }
      token.clear();
      continue;
    }
    token += ch;
  }

  elems.push_back(token);
}

static void RemoveEmptyTokens(std::vector<std::string> *tokens) {
  if (!tokens) return;

  const std::vector<std::string> &src = *tokens;
  std::vector<std::string> filtered;
  filtered.reserve(src.size());
  for (size_t i = 0; i < src.size(); i++) {
    if (!src[i].empty()) {
      filtered.push_back(src[i]);
    }
  }
  tokens->swap(filtered);
}

static std::string JoinPath(const std::string &dir,
                            const std::string &filename) {
  if (dir.empty()) {
    return filename;
  } else {
    // check '/'
    char lastChar = *dir.rbegin();
    if (lastChar != '/') {
      return dir + std::string("/") + filename;
    } else {
      return dir + filename;
    }
  }
}

static bool LoadMtlInternal(std::map<std::string, int> *material_map,
                            std::vector<material_t> *materials,
                            StreamReader &sr,
                            std::string *warning, std::string *err,
                            const std::string &filename = "<stream>") {
  if (sr.has_errors()) {
    if (err) {
      (*err) += sr.get_errors();
    }
    return false;
  }

  material_t material;
  InitMaterial(&material);

  // Issue 43. `d` wins against `Tr` since `Tr` is not in the MTL specification.
  bool has_d = false;
  bool has_tr = false;

  // has_kd is used to set a default diffuse value when map_Kd is present
  // and Kd is not.
  bool has_kd = false;

  std::stringstream warn_ss;

  // Handle BOM
  if (sr.remaining() >= 3 &&
      static_cast<unsigned char>(sr.peek()) == 0xEF &&
      static_cast<unsigned char>(sr.peek_at(1)) == 0xBB &&
      static_cast<unsigned char>(sr.peek_at(2)) == 0xBF) {
    sr.advance(3);
  }

  while (!sr.eof()) {
    sr.skip_space();
    if (sr.at_line_end()) { sr.skip_line(); continue; }
    if (sr.peek() == '#') { sr.skip_line(); continue; }

    size_t line_num = sr.line_num();

    // new mtl
    if (sr.match("newmtl", 6) && (sr.peek_at(6) == ' ' || sr.peek_at(6) == '\t')) {
      // flush previous material.
      if (!material.name.empty()) {
        material_map->insert(std::pair<std::string, int>(
            material.name, static_cast<int>(materials->size())));
        materials->push_back(material);
      }

      InitMaterial(&material);

      has_d = false;
      has_tr = false;
      has_kd = false;

      sr.advance(7);
      {
        std::string namebuf = sr_parseString(sr);
        if (namebuf.empty()) {
          if (warning) {
            (*warning) += "empty material name in `newmtl`\n";
          }
        }
        material.name = namebuf;
      }
      sr.skip_line();
      continue;
    }

    // ambient
    if (sr.peek() == 'K' && sr.peek_at(1) == 'a' && (sr.peek_at(2) == ' ' || sr.peek_at(2) == '\t')) {
      sr.advance(2);
      real_t r, g, b;
      if (!sr_parseReal3(&r, &g, &b, sr, err, filename)) return false;
      material.ambient[0] = r;
      material.ambient[1] = g;
      material.ambient[2] = b;
      sr.skip_line();
      continue;
    }

    // diffuse
    if (sr.peek() == 'K' && sr.peek_at(1) == 'd' && (sr.peek_at(2) == ' ' || sr.peek_at(2) == '\t')) {
      sr.advance(2);
      real_t r, g, b;
      if (!sr_parseReal3(&r, &g, &b, sr, err, filename)) return false;
      material.diffuse[0] = r;
      material.diffuse[1] = g;
      material.diffuse[2] = b;
      has_kd = true;
      sr.skip_line();
      continue;
    }

    // specular
    if (sr.peek() == 'K' && sr.peek_at(1) == 's' && (sr.peek_at(2) == ' ' || sr.peek_at(2) == '\t')) {
      sr.advance(2);
      real_t r, g, b;
      if (!sr_parseReal3(&r, &g, &b, sr, err, filename)) return false;
      material.specular[0] = r;
      material.specular[1] = g;
      material.specular[2] = b;
      sr.skip_line();
      continue;
    }

    // transmittance
    if ((sr.peek() == 'K' && sr.peek_at(1) == 't' && (sr.peek_at(2) == ' ' || sr.peek_at(2) == '\t')) ||
        (sr.peek() == 'T' && sr.peek_at(1) == 'f' && (sr.peek_at(2) == ' ' || sr.peek_at(2) == '\t'))) {
      sr.advance(2);
      real_t r, g, b;
      if (!sr_parseReal3(&r, &g, &b, sr, err, filename)) return false;
      material.transmittance[0] = r;
      material.transmittance[1] = g;
      material.transmittance[2] = b;
      sr.skip_line();
      continue;
    }

    // ior(index of refraction)
    if (sr.peek() == 'N' && sr.peek_at(1) == 'i' && (sr.peek_at(2) == ' ' || sr.peek_at(2) == '\t')) {
      sr.advance(2);
      if (!sr_parseReal(sr, &material.ior, 0.0, err, filename)) return false;
      sr.skip_line();
      continue;
    }

    // emission
    if (sr.peek() == 'K' && sr.peek_at(1) == 'e' && (sr.peek_at(2) == ' ' || sr.peek_at(2) == '\t')) {
      sr.advance(2);
      real_t r, g, b;
      if (!sr_parseReal3(&r, &g, &b, sr, err, filename)) return false;
      material.emission[0] = r;
      material.emission[1] = g;
      material.emission[2] = b;
      sr.skip_line();
      continue;
    }

    // shininess
    if (sr.peek() == 'N' && sr.peek_at(1) == 's' && (sr.peek_at(2) == ' ' || sr.peek_at(2) == '\t')) {
      sr.advance(2);
      if (!sr_parseReal(sr, &material.shininess, 0.0, err, filename)) return false;
      sr.skip_line();
      continue;
    }

    // illum model
    if (sr.match("illum", 5) && (sr.peek_at(5) == ' ' || sr.peek_at(5) == '\t')) {
      sr.advance(6);
      if (!sr_parseInt(sr, &material.illum, err, filename)) return false;
      sr.skip_line();
      continue;
    }

    // dissolve
    if (sr.peek() == 'd' && (sr.peek_at(1) == ' ' || sr.peek_at(1) == '\t')) {
      sr.advance(1);
      if (!sr_parseReal(sr, &material.dissolve, 0.0, err, filename)) return false;

      if (has_tr) {
        warn_ss << "Both `d` and `Tr` parameters defined for \""
                << material.name
                << "\". Use the value of `d` for dissolve (line " << line_num
                << " in .mtl.)\n";
      }
      has_d = true;
      sr.skip_line();
      continue;
    }
    if (sr.peek() == 'T' && sr.peek_at(1) == 'r' && (sr.peek_at(2) == ' ' || sr.peek_at(2) == '\t')) {
      sr.advance(2);
      if (has_d) {
        warn_ss << "Both `d` and `Tr` parameters defined for \""
                << material.name
                << "\". Use the value of `d` for dissolve (line " << line_num
                << " in .mtl.)\n";
      } else {
        real_t tr_val;
        if (!sr_parseReal(sr, &tr_val, 0.0, err, filename)) return false;
        material.dissolve = static_cast<real_t>(1.0) - tr_val;
      }
      has_tr = true;
      sr.skip_line();
      continue;
    }

    // PBR: roughness
    if (sr.peek() == 'P' && sr.peek_at(1) == 'r' && (sr.peek_at(2) == ' ' || sr.peek_at(2) == '\t')) {
      sr.advance(2);
      if (!sr_parseReal(sr, &material.roughness, 0.0, err, filename)) return false;
      sr.skip_line();
      continue;
    }

    // PBR: metallic
    if (sr.peek() == 'P' && sr.peek_at(1) == 'm' && (sr.peek_at(2) == ' ' || sr.peek_at(2) == '\t')) {
      sr.advance(2);
      if (!sr_parseReal(sr, &material.metallic, 0.0, err, filename)) return false;
      sr.skip_line();
      continue;
    }

    // PBR: sheen
    if (sr.peek() == 'P' && sr.peek_at(1) == 's' && (sr.peek_at(2) == ' ' || sr.peek_at(2) == '\t')) {
      sr.advance(2);
      if (!sr_parseReal(sr, &material.sheen, 0.0, err, filename)) return false;
      sr.skip_line();
      continue;
    }

    // PBR: clearcoat thickness
    if (sr.peek() == 'P' && sr.peek_at(1) == 'c' && (sr.peek_at(2) == ' ' || sr.peek_at(2) == '\t')) {
      sr.advance(2);
      if (!sr_parseReal(sr, &material.clearcoat_thickness, 0.0, err, filename)) return false;
      sr.skip_line();
      continue;
    }

    // PBR: clearcoat roughness
    if (sr.match("Pcr", 3) && (sr.peek_at(3) == ' ' || sr.peek_at(3) == '\t')) {
      sr.advance(4);
      if (!sr_parseReal(sr, &material.clearcoat_roughness, 0.0, err, filename)) return false;
      sr.skip_line();
      continue;
    }

    // PBR: anisotropy
    if (sr.match("aniso", 5) && (sr.peek_at(5) == ' ' || sr.peek_at(5) == '\t')) {
      sr.advance(6);
      if (!sr_parseReal(sr, &material.anisotropy, 0.0, err, filename)) return false;
      sr.skip_line();
      continue;
    }

    // PBR: anisotropy rotation
    if (sr.match("anisor", 6) && (sr.peek_at(6) == ' ' || sr.peek_at(6) == '\t')) {
      sr.advance(7);
      if (!sr_parseReal(sr, &material.anisotropy_rotation, 0.0, err, filename)) return false;
      sr.skip_line();
      continue;
    }

    // For texture directives, read rest of line and delegate to
    // ParseTextureNameAndOption (which uses the old const char* parse functions).

    // ambient or ambient occlusion texture
    if (sr.match("map_Ka", 6) && (sr.peek_at(6) == ' ' || sr.peek_at(6) == '\t')) {
      sr.advance(7);
      std::string line_rest = trimTrailingWhitespace(sr.read_line());
      ParseTextureNameAndOption(&(material.ambient_texname),
                                &(material.ambient_texopt), line_rest.c_str());
      sr.skip_line();
      continue;
    }

    // diffuse texture
    if (sr.match("map_Kd", 6) && (sr.peek_at(6) == ' ' || sr.peek_at(6) == '\t')) {
      sr.advance(7);
      std::string line_rest = trimTrailingWhitespace(sr.read_line());
      ParseTextureNameAndOption(&(material.diffuse_texname),
                                &(material.diffuse_texopt), line_rest.c_str());
      if (!has_kd) {
        material.diffuse[0] = static_cast<real_t>(0.6);
        material.diffuse[1] = static_cast<real_t>(0.6);
        material.diffuse[2] = static_cast<real_t>(0.6);
      }
      sr.skip_line();
      continue;
    }

    // specular texture
    if (sr.match("map_Ks", 6) && (sr.peek_at(6) == ' ' || sr.peek_at(6) == '\t')) {
      sr.advance(7);
      std::string line_rest = trimTrailingWhitespace(sr.read_line());
      ParseTextureNameAndOption(&(material.specular_texname),
                                &(material.specular_texopt), line_rest.c_str());
      sr.skip_line();
      continue;
    }

    // specular highlight texture
    if (sr.match("map_Ns", 6) && (sr.peek_at(6) == ' ' || sr.peek_at(6) == '\t')) {
      sr.advance(7);
      std::string line_rest = trimTrailingWhitespace(sr.read_line());
      ParseTextureNameAndOption(&(material.specular_highlight_texname),
                                &(material.specular_highlight_texopt), line_rest.c_str());
      sr.skip_line();
      continue;
    }

    // bump texture
    if ((sr.match("map_bump", 8) || sr.match("map_Bump", 8)) &&
        (sr.peek_at(8) == ' ' || sr.peek_at(8) == '\t')) {
      sr.advance(9);
      std::string line_rest = trimTrailingWhitespace(sr.read_line());
      ParseTextureNameAndOption(&(material.bump_texname),
                                &(material.bump_texopt), line_rest.c_str());
      sr.skip_line();
      continue;
    }

    // bump texture (short form)
    if (sr.match("bump", 4) && (sr.peek_at(4) == ' ' || sr.peek_at(4) == '\t')) {
      sr.advance(5);
      std::string line_rest = trimTrailingWhitespace(sr.read_line());
      ParseTextureNameAndOption(&(material.bump_texname),
                                &(material.bump_texopt), line_rest.c_str());
      sr.skip_line();
      continue;
    }

    // alpha texture
    if (sr.match("map_d", 5) && (sr.peek_at(5) == ' ' || sr.peek_at(5) == '\t')) {
      sr.advance(6);
      std::string line_rest = trimTrailingWhitespace(sr.read_line());
      ParseTextureNameAndOption(&(material.alpha_texname),
                                &(material.alpha_texopt), line_rest.c_str());
      sr.skip_line();
      continue;
    }

    // displacement texture
    if ((sr.match("map_disp", 8) || sr.match("map_Disp", 8)) &&
        (sr.peek_at(8) == ' ' || sr.peek_at(8) == '\t')) {
      sr.advance(9);
      std::string line_rest = trimTrailingWhitespace(sr.read_line());
      ParseTextureNameAndOption(&(material.displacement_texname),
                                &(material.displacement_texopt), line_rest.c_str());
      sr.skip_line();
      continue;
    }

    // displacement texture (short form)
    if (sr.match("disp", 4) && (sr.peek_at(4) == ' ' || sr.peek_at(4) == '\t')) {
      sr.advance(5);
      std::string line_rest = trimTrailingWhitespace(sr.read_line());
      ParseTextureNameAndOption(&(material.displacement_texname),
                                &(material.displacement_texopt), line_rest.c_str());
      sr.skip_line();
      continue;
    }

    // reflection map
    if (sr.match("refl", 4) && (sr.peek_at(4) == ' ' || sr.peek_at(4) == '\t')) {
      sr.advance(5);
      std::string line_rest = trimTrailingWhitespace(sr.read_line());
      ParseTextureNameAndOption(&(material.reflection_texname),
                                &(material.reflection_texopt), line_rest.c_str());
      sr.skip_line();
      continue;
    }

    // PBR: roughness texture
    if (sr.match("map_Pr", 6) && (sr.peek_at(6) == ' ' || sr.peek_at(6) == '\t')) {
      sr.advance(7);
      std::string line_rest = trimTrailingWhitespace(sr.read_line());
      ParseTextureNameAndOption(&(material.roughness_texname),
                                &(material.roughness_texopt), line_rest.c_str());
      sr.skip_line();
      continue;
    }

    // PBR: metallic texture
    if (sr.match("map_Pm", 6) && (sr.peek_at(6) == ' ' || sr.peek_at(6) == '\t')) {
      sr.advance(7);
      std::string line_rest = trimTrailingWhitespace(sr.read_line());
      ParseTextureNameAndOption(&(material.metallic_texname),
                                &(material.metallic_texopt), line_rest.c_str());
      sr.skip_line();
      continue;
    }

    // PBR: sheen texture
    if (sr.match("map_Ps", 6) && (sr.peek_at(6) == ' ' || sr.peek_at(6) == '\t')) {
      sr.advance(7);
      std::string line_rest = trimTrailingWhitespace(sr.read_line());
      ParseTextureNameAndOption(&(material.sheen_texname),
                                &(material.sheen_texopt), line_rest.c_str());
      sr.skip_line();
      continue;
    }

    // PBR: emissive texture
    if (sr.match("map_Ke", 6) && (sr.peek_at(6) == ' ' || sr.peek_at(6) == '\t')) {
      sr.advance(7);
      std::string line_rest = trimTrailingWhitespace(sr.read_line());
      ParseTextureNameAndOption(&(material.emissive_texname),
                                &(material.emissive_texopt), line_rest.c_str());
      sr.skip_line();
      continue;
    }

    // PBR: normal map texture
    if (sr.match("norm", 4) && (sr.peek_at(4) == ' ' || sr.peek_at(4) == '\t')) {
      sr.advance(5);
      std::string line_rest = trimTrailingWhitespace(sr.read_line());
      ParseTextureNameAndOption(&(material.normal_texname),
                                &(material.normal_texopt), line_rest.c_str());
      sr.skip_line();
      continue;
    }

    // unknown parameter
    {
      std::string line_rest = trimTrailingWhitespace(sr.read_line());
      const char *_lp = line_rest.c_str();
      const char *_space = strchr(_lp, ' ');
      if (!_space) {
        _space = strchr(_lp, '\t');
      }
      if (_space) {
        std::ptrdiff_t len = _space - _lp;
        std::string key(_lp, static_cast<size_t>(len));
        std::string value = _space + 1;
        material.unknown_parameter.insert(
            std::pair<std::string, std::string>(key, value));
      }
    }
    sr.skip_line();
  }
  // flush last material (only if it was actually defined).
  if (!material.name.empty()) {
    material_map->insert(std::pair<std::string, int>(
        material.name, static_cast<int>(materials->size())));
    materials->push_back(material);
  }

  if (warning) {
    (*warning) += warn_ss.str();
  }

  return true;
}

void LoadMtl(std::map<std::string, int> *material_map,
             std::vector<material_t> *materials, std::istream *inStream,
             std::string *warning, std::string *err) {
  StreamReader sr(*inStream);
  LoadMtlInternal(material_map, materials, sr, warning, err);
}


bool MaterialFileReader::operator()(const std::string &matId,
                                    std::vector<material_t> *materials,
                                    std::map<std::string, int> *matMap,
                                    std::string *warn, std::string *err) {
  if (!m_mtlBaseDir.empty()) {
#ifdef _WIN32
    char sep = ';';
#else
    char sep = ':';
#endif

    // https://stackoverflow.com/questions/5167625/splitting-a-c-stdstring-using-tokens-e-g
    std::vector<std::string> paths;
    std::istringstream f(m_mtlBaseDir);

    std::string s;
    while (getline(f, s, sep)) {
      paths.push_back(s);
    }

    for (size_t i = 0; i < paths.size(); i++) {
      std::string filepath = JoinPath(paths[i], matId);

#ifdef TINYOBJLOADER_USE_MMAP
      {
        MappedFile mf;
        if (!mf.open(filepath.c_str())) continue;
        if (mf.size > TINYOBJLOADER_STREAM_READER_MAX_BYTES) {
          if (err) {
            std::stringstream ss;
            ss << "input stream too large (" << mf.size
               << " bytes exceeds limit "
               << TINYOBJLOADER_STREAM_READER_MAX_BYTES << " bytes)\n";
            (*err) += ss.str();
          }
          return false;
        }
        StreamReader sr(mf.data, mf.size);
        return LoadMtlInternal(matMap, materials, sr, warn, err, filepath);
      }
#else   // !TINYOBJLOADER_USE_MMAP
#ifdef _WIN32
      std::ifstream matIStream(LongPathW(UTF8ToWchar(filepath)).c_str());
#else
      std::ifstream matIStream(filepath.c_str());
#endif
      if (matIStream) {
        StreamReader mtl_sr(matIStream);
        return LoadMtlInternal(matMap, materials, mtl_sr, warn, err, filepath);
      }
#endif  // TINYOBJLOADER_USE_MMAP
    }

    std::stringstream ss;
    ss << "Material file [ " << matId
       << " ] not found in a path : " << m_mtlBaseDir << "\n";
    if (warn) {
      (*warn) += ss.str();
    }
    return false;

  } else {
    std::string filepath = matId;

#ifdef TINYOBJLOADER_USE_MMAP
    {
      MappedFile mf;
      if (mf.open(filepath.c_str())) {
        if (mf.size > TINYOBJLOADER_STREAM_READER_MAX_BYTES) {
          if (err) {
            std::stringstream ss;
            ss << "input stream too large (" << mf.size
               << " bytes exceeds limit "
               << TINYOBJLOADER_STREAM_READER_MAX_BYTES << " bytes)\n";
            (*err) += ss.str();
          }
          return false;
        }
        StreamReader sr(mf.data, mf.size);
        return LoadMtlInternal(matMap, materials, sr, warn, err, filepath);
      }
    }
#else   // !TINYOBJLOADER_USE_MMAP
#ifdef _WIN32
    std::ifstream matIStream(LongPathW(UTF8ToWchar(filepath)).c_str());
#else
    std::ifstream matIStream(filepath.c_str());
#endif
    if (matIStream) {
      StreamReader mtl_sr(matIStream);
      return LoadMtlInternal(matMap, materials, mtl_sr, warn, err, filepath);
    }
#endif  // TINYOBJLOADER_USE_MMAP

    std::stringstream ss;
    ss << "Material file [ " << filepath
       << " ] not found in a path : " << m_mtlBaseDir << "\n";
    if (warn) {
      (*warn) += ss.str();
    }

    return false;
  }
}

bool MaterialStreamReader::operator()(const std::string &matId,
                                      std::vector<material_t> *materials,
                                      std::map<std::string, int> *matMap,
                                      std::string *warn, std::string *err) {
  (void)matId;
  if (!m_inStream) {
    std::stringstream ss;
    ss << "Material stream in error state. \n";
    if (warn) {
      (*warn) += ss.str();
    }
    return false;
  }

  StreamReader mtl_sr(m_inStream);
  return LoadMtlInternal(matMap, materials, mtl_sr, warn, err, "<stream>");
}

static bool LoadObjInternal(attrib_t *attrib, std::vector<shape_t> *shapes,
                            std::vector<material_t> *materials,
                            std::string *warn, std::string *err,
                            StreamReader &sr,
                            MaterialReader *readMatFn, bool triangulate,
                            bool default_vcols_fallback,
                            const std::string &filename = "<stream>") {
  if (sr.has_errors()) {
    if (err) {
      (*err) += sr.get_errors();
    }
    return false;
  }

  std::vector<real_t> v;
  std::vector<real_t> vertex_weights;
  std::vector<real_t> vn;
  std::vector<real_t> vt;
  std::vector<real_t> vt_w;  // optional [w] component in `vt`
  std::vector<real_t> vc;
  std::vector<skin_weight_t> vw;
  std::vector<tag_t> tags;
  PrimGroup prim_group;
  std::string name;

  // material
  std::set<std::string> material_filenames;
  std::map<std::string, int> material_map;
  int material = -1;

  unsigned int current_smoothing_id = 0;

  int greatest_v_idx = -1;
  int greatest_vn_idx = -1;
  int greatest_vt_idx = -1;

  shape_t shape;

  bool found_all_colors = true;

  // Handle BOM
  if (sr.remaining() >= 3 &&
      static_cast<unsigned char>(sr.peek()) == 0xEF &&
      static_cast<unsigned char>(sr.peek_at(1)) == 0xBB &&
      static_cast<unsigned char>(sr.peek_at(2)) == 0xBF) {
    sr.advance(3);
  }

  warning_context context;
  context.warn = warn;
  context.filename = filename;

  while (!sr.eof()) {
    sr.skip_space();
    if (sr.at_line_end()) { sr.skip_line(); continue; }
    if (sr.peek() == '#') { sr.skip_line(); continue; }

    size_t line_num = sr.line_num();

    // vertex
    if (sr.peek() == 'v' && (sr.peek_at(1) == ' ' || sr.peek_at(1) == '\t')) {
      sr.advance(2);
      real_t x, y, z;
      real_t r, g, b;

      int num_components = sr_parseVertexWithColor(&x, &y, &z, &r, &g, &b, sr, err, filename);
      if (num_components < 0) return false;
      found_all_colors &= (num_components == 6);

      v.push_back(x);
      v.push_back(y);
      v.push_back(z);

      vertex_weights.push_back(r);

      if ((num_components == 6) || default_vcols_fallback) {
        vc.push_back(r);
        vc.push_back(g);
        vc.push_back(b);
      }

      sr.skip_line();
      continue;
    }

    // normal
    if (sr.peek() == 'v' && sr.peek_at(1) == 'n' && (sr.peek_at(2) == ' ' || sr.peek_at(2) == '\t')) {
      sr.advance(3);
      real_t x, y, z;
      if (!sr_parseReal3(&x, &y, &z, sr, err, filename)) return false;
      vn.push_back(x);
      vn.push_back(y);
      vn.push_back(z);
      sr.skip_line();
      continue;
    }

    // texcoord
    if (sr.peek() == 'v' && sr.peek_at(1) == 't' && (sr.peek_at(2) == ' ' || sr.peek_at(2) == '\t')) {
      sr.advance(3);
      real_t x, y;
      if (!sr_parseReal2(&x, &y, sr, err, filename)) return false;
      vt.push_back(x);
      vt.push_back(y);

      // Parse optional w component
      real_t w = static_cast<real_t>(0.0);
      sr_parseReal(sr, &w);
      vt_w.push_back(w);

      sr.skip_line();
      continue;
    }

    // skin weight. tinyobj extension
    if (sr.peek() == 'v' && sr.peek_at(1) == 'w' && (sr.peek_at(2) == ' ' || sr.peek_at(2) == '\t')) {
      sr.advance(3);

      int vid;
      if (!sr_parseInt(sr, &vid, err, filename)) return false;

      skin_weight_t sw;
      sw.vertex_id = vid;

      size_t vw_loop_max = sr.remaining() + 1;
      size_t vw_loop_iter = 0;
      while (!sr.at_line_end() && sr.peek() != '#' &&
             vw_loop_iter < vw_loop_max) {
        real_t j, w;
        sr_parseReal2(&j, &w, sr, -1.0);

        if (j < static_cast<real_t>(0)) {
          if (err) {
            (*err) += sr.format_error(filename,
                "failed to parse `vw' line: joint_id is negative");
          }
          return false;
        }

        // Clamp to int range to avoid UB on float-to-int overflow.
        if (j > static_cast<real_t>(std::numeric_limits<int>::max())) {
          if (err) {
            (*err) += sr.format_error(filename,
                "failed to parse `vw' line: joint_id overflow");
          }
          return false;
        }
        joint_and_weight_t jw;
        jw.joint_id = static_cast<int>(j);
        jw.weight = w;

        sw.weightValues.push_back(jw);
        sr.skip_space_and_cr();
        vw_loop_iter++;
      }

      vw.push_back(sw);
      sr.skip_line();
      continue;
    }

    context.line_number = line_num;

    // line
    if (sr.peek() == 'l' && (sr.peek_at(1) == ' ' || sr.peek_at(1) == '\t')) {
      sr.advance(2);

      __line_t line;

      size_t l_loop_max = sr.remaining() + 1;
      size_t l_loop_iter = 0;
      while (!sr.at_line_end() && sr.peek() != '#' &&
             l_loop_iter < l_loop_max) {
        vertex_index_t vi;
        if (!sr_parseTriple(sr, size_to_int(v.size() / 3),
                         size_to_int(vn.size() / 3),
                         size_to_int(vt.size() / 2), &vi, context)) {
          if (err) {
            (*err) += sr.format_error(filename,
                "failed to parse `l' line (invalid vertex index)");
          }
          return false;
        }

        line.vertex_indices.push_back(vi);
        sr.skip_space_and_cr();
        l_loop_iter++;
      }

      prim_group.lineGroup.push_back(line);
      sr.skip_line();
      continue;
    }

    // points
    if (sr.peek() == 'p' && (sr.peek_at(1) == ' ' || sr.peek_at(1) == '\t')) {
      sr.advance(2);

      __points_t pts;

      size_t p_loop_max = sr.remaining() + 1;
      size_t p_loop_iter = 0;
      while (!sr.at_line_end() && sr.peek() != '#' &&
             p_loop_iter < p_loop_max) {
        vertex_index_t vi;
        if (!sr_parseTriple(sr, size_to_int(v.size() / 3),
                         size_to_int(vn.size() / 3),
                         size_to_int(vt.size() / 2), &vi, context)) {
          if (err) {
            (*err) += sr.format_error(filename,
                "failed to parse `p' line (invalid vertex index)");
          }
          return false;
        }

        pts.vertex_indices.push_back(vi);
        sr.skip_space_and_cr();
        p_loop_iter++;
      }

      prim_group.pointsGroup.push_back(pts);
      sr.skip_line();
      continue;
    }

    // face
    if (sr.peek() == 'f' && (sr.peek_at(1) == ' ' || sr.peek_at(1) == '\t')) {
      sr.advance(2);
      sr.skip_space();

      face_t face;

      face.smoothing_group_id = current_smoothing_id;
      face.vertex_indices.reserve(3);

      size_t f_loop_max = sr.remaining() + 1;
      size_t f_loop_iter = 0;
      while (!sr.at_line_end() && sr.peek() != '#' &&
             f_loop_iter < f_loop_max) {
        vertex_index_t vi;
        if (!sr_parseTriple(sr, size_to_int(v.size() / 3),
                         size_to_int(vn.size() / 3),
                         size_to_int(vt.size() / 2), &vi, context)) {
          if (err) {
            (*err) += sr.format_error(filename,
                "failed to parse `f' line (invalid vertex index)");
          }
          return false;
        }

        greatest_v_idx = greatest_v_idx > vi.v_idx ? greatest_v_idx : vi.v_idx;
        greatest_vn_idx =
            greatest_vn_idx > vi.vn_idx ? greatest_vn_idx : vi.vn_idx;
        greatest_vt_idx =
            greatest_vt_idx > vi.vt_idx ? greatest_vt_idx : vi.vt_idx;

        face.vertex_indices.push_back(vi);
        sr.skip_space_and_cr();
        f_loop_iter++;
      }

      prim_group.faceGroup.push_back(face);
      sr.skip_line();
      continue;
    }

    // use mtl
    if (sr.match("usemtl", 6) && (sr.peek_at(6) == ' ' || sr.peek_at(6) == '\t')) {
      sr.advance(6);
      std::string namebuf = sr_parseString(sr);

      int newMaterialId = -1;
      std::map<std::string, int>::const_iterator it =
          material_map.find(namebuf);
      if (it != material_map.end()) {
        newMaterialId = it->second;
      } else {
        if (warn) {
          (*warn) += "material [ '" + namebuf + "' ] not found in .mtl\n";
        }
      }

      if (newMaterialId != material) {
        exportGroupsToShape(&shape, prim_group, tags, material, name,
                            triangulate, v, warn);
        prim_group.faceGroup.clear();
        material = newMaterialId;
      }

      sr.skip_line();
      continue;
    }

    // load mtl
    if (sr.match("mtllib", 6) && (sr.peek_at(6) == ' ' || sr.peek_at(6) == '\t')) {
      if (readMatFn) {
        sr.advance(7);

        std::string line_rest = trimTrailingWhitespace(sr.read_line());
        std::vector<std::string> filenames;
        SplitString(line_rest, ' ', '\\', filenames);
        RemoveEmptyTokens(&filenames);

        if (filenames.empty()) {
          if (warn) {
            std::stringstream ss;
            ss << "Looks like empty filename for mtllib. Use default "
                  "material (line "
               << line_num << ".)\n";

            (*warn) += ss.str();
          }
        } else {
          bool found = false;
          for (size_t s = 0; s < filenames.size(); s++) {
            if (material_filenames.count(filenames[s]) > 0) {
              found = true;
              continue;
            }

            std::string warn_mtl;
            std::string err_mtl;
            bool ok = (*readMatFn)(filenames[s].c_str(), materials,
                                   &material_map, &warn_mtl, &err_mtl);
            if (warn && (!warn_mtl.empty())) {
              (*warn) += warn_mtl;
            }

            if (err && (!err_mtl.empty())) {
              (*err) += err_mtl;
            }

            if (ok) {
              found = true;
              material_filenames.insert(filenames[s]);
              break;
            }
          }

          if (!found) {
            if (warn) {
              (*warn) +=
                  "Failed to load material file(s). Use default "
                  "material.\n";
            }
          }
        }
      }

      sr.skip_line();
      continue;
    }

    // group name
    if (sr.peek() == 'g' && (sr.peek_at(1) == ' ' || sr.peek_at(1) == '\t')) {
      // flush previous face group.
      bool ret = exportGroupsToShape(&shape, prim_group, tags, material, name,
                                     triangulate, v, warn);
      (void)ret;

      if (shape.mesh.indices.size() > 0) {
        shapes->push_back(shape);
      }

      shape = shape_t();

      // material = -1;
      prim_group.clear();

      std::vector<std::string> names;

      size_t g_loop_max = sr.remaining() + 1;
      size_t g_loop_iter = 0;
      while (!sr.at_line_end() && sr.peek() != '#' &&
             g_loop_iter < g_loop_max) {
        std::string str = sr_parseString(sr);
        names.push_back(str);
        sr.skip_space_and_cr();
        g_loop_iter++;
      }

      // names[0] must be 'g'

      if (names.size() < 2) {
        // 'g' with empty names
        if (warn) {
          std::stringstream ss;
          ss << "Empty group name. line: " << line_num << "\n";
          (*warn) += ss.str();
          name = "";
        }
      } else {
        std::stringstream ss;
        ss << names[1];

        for (size_t i = 2; i < names.size(); i++) {
          ss << " " << names[i];
        }

        name = ss.str();
      }

      sr.skip_line();
      continue;
    }

    // object name
    if (sr.peek() == 'o' && (sr.peek_at(1) == ' ' || sr.peek_at(1) == '\t')) {
      // flush previous face group.
      bool ret = exportGroupsToShape(&shape, prim_group, tags, material, name,
                                     triangulate, v, warn);
      (void)ret;

      if (shape.mesh.indices.size() > 0 || shape.lines.indices.size() > 0 ||
          shape.points.indices.size() > 0) {
        shapes->push_back(shape);
      }

      // material = -1;
      prim_group.clear();
      shape = shape_t();

      sr.advance(2);
      std::string rest = sr.read_line();
      name = rest;

      sr.skip_line();
      continue;
    }

    if (sr.peek() == 't' && (sr.peek_at(1) == ' ' || sr.peek_at(1) == '\t')) {
      const int max_tag_nums = 8192;
      tag_t tag;

      sr.advance(2);

      tag.name = sr_parseString(sr);

      tag_sizes ts = sr_parseTagTriple(sr);

      if (ts.num_ints < 0) {
        ts.num_ints = 0;
      }
      if (ts.num_ints > max_tag_nums) {
        ts.num_ints = max_tag_nums;
      }

      if (ts.num_reals < 0) {
        ts.num_reals = 0;
      }
      if (ts.num_reals > max_tag_nums) {
        ts.num_reals = max_tag_nums;
      }

      if (ts.num_strings < 0) {
        ts.num_strings = 0;
      }
      if (ts.num_strings > max_tag_nums) {
        ts.num_strings = max_tag_nums;
      }

      tag.intValues.resize(static_cast<size_t>(ts.num_ints));

      for (size_t i = 0; i < static_cast<size_t>(ts.num_ints); ++i) {
        tag.intValues[i] = sr_parseInt(sr);
      }

      tag.floatValues.resize(static_cast<size_t>(ts.num_reals));
      for (size_t i = 0; i < static_cast<size_t>(ts.num_reals); ++i) {
        tag.floatValues[i] = sr_parseReal(sr);
      }

      tag.stringValues.resize(static_cast<size_t>(ts.num_strings));
      for (size_t i = 0; i < static_cast<size_t>(ts.num_strings); ++i) {
        tag.stringValues[i] = sr_parseString(sr);
      }

      tags.push_back(tag);

      sr.skip_line();
      continue;
    }

    if (sr.peek() == 's' && (sr.peek_at(1) == ' ' || sr.peek_at(1) == '\t')) {
      // smoothing group id
      sr.advance(2);
      sr.skip_space();

      if (sr.at_line_end()) {
        sr.skip_line();
        continue;
      }

      if (sr.peek() == '\r') {
        sr.skip_line();
        continue;
      }

      if (sr.remaining() >= 3 && sr.match("off", 3)) {
        current_smoothing_id = 0;
      } else {
        int smGroupId = sr_parseInt(sr);
        if (smGroupId < 0) {
          current_smoothing_id = 0;
        } else {
          current_smoothing_id = static_cast<unsigned int>(smGroupId);
        }
      }

      sr.skip_line();
      continue;
    }

    // Ignore unknown command.
    sr.skip_line();
  }

  // not all vertices have colors, no default colors desired? -> clear colors
  if (!found_all_colors && !default_vcols_fallback) {
    vc.clear();
  }

  if (greatest_v_idx >= size_to_int(v.size() / 3)) {
    if (warn) {
      std::stringstream ss;
      ss << "Vertex indices out of bounds (line " << sr.line_num() << ".)\n\n";
      (*warn) += ss.str();
    }
  }
  if (greatest_vn_idx >= size_to_int(vn.size() / 3)) {
    if (warn) {
      std::stringstream ss;
      ss << "Vertex normal indices out of bounds (line " << sr.line_num()
         << ".)\n\n";
      (*warn) += ss.str();
    }
  }
  if (greatest_vt_idx >= size_to_int(vt.size() / 2)) {
    if (warn) {
      std::stringstream ss;
      ss << "Vertex texcoord indices out of bounds (line " << sr.line_num()
         << ".)\n\n";
      (*warn) += ss.str();
    }
  }

  bool ret = exportGroupsToShape(&shape, prim_group, tags, material, name,
                                 triangulate, v, warn);
  if (ret || shape.mesh.indices.size()) {
    shapes->push_back(shape);
  }
  prim_group.clear();

  attrib->vertices.swap(v);
  attrib->vertex_weights.swap(vertex_weights);
  attrib->normals.swap(vn);
  attrib->texcoords.swap(vt);
  attrib->texcoord_ws.swap(vt_w);
  attrib->colors.swap(vc);
  attrib->skin_weights.swap(vw);

  return true;
}

bool LoadObj(attrib_t *attrib, std::vector<shape_t> *shapes,
             std::vector<material_t> *materials, std::string *warn,
             std::string *err, const char *filename, const char *mtl_basedir,
             bool triangulate, bool default_vcols_fallback) {
  attrib->vertices.clear();
  attrib->vertex_weights.clear();
  attrib->normals.clear();
  attrib->texcoords.clear();
  attrib->texcoord_ws.clear();
  attrib->colors.clear();
  attrib->skin_weights.clear();
  shapes->clear();

  std::string baseDir = mtl_basedir ? mtl_basedir : "";
  if (!baseDir.empty()) {
#ifndef _WIN32
    const char dirsep = '/';
#else
    const char dirsep = '\\';
#endif
    if (baseDir[baseDir.length() - 1] != dirsep) baseDir += dirsep;
  }
  MaterialFileReader matFileReader(baseDir);

#ifdef TINYOBJLOADER_USE_MMAP
  {
    MappedFile mf;
    if (!mf.open(filename)) {
      if (err) {
        std::stringstream ss;
        ss << "Cannot open file [" << filename << "]\n";
        (*err) = ss.str();
      }
      return false;
    }
    if (mf.size > TINYOBJLOADER_STREAM_READER_MAX_BYTES) {
      if (err) {
        std::stringstream ss;
        ss << "input stream too large (" << mf.size
           << " bytes exceeds limit "
           << TINYOBJLOADER_STREAM_READER_MAX_BYTES << " bytes)\n";
        (*err) += ss.str();
      }
      return false;
    }
    StreamReader sr(mf.data, mf.size);
    return LoadObjInternal(attrib, shapes, materials, warn, err, sr,
                           &matFileReader, triangulate, default_vcols_fallback,
                           filename);
  }
#else   // !TINYOBJLOADER_USE_MMAP
#ifdef _WIN32
  std::ifstream ifs(LongPathW(UTF8ToWchar(filename)).c_str());
#else
  std::ifstream ifs(filename);
#endif
  if (!ifs) {
    if (err) {
      std::stringstream ss;
      ss << "Cannot open file [" << filename << "]\n";
      (*err) = ss.str();
    }
    return false;
  }
  {
    StreamReader sr(ifs);
    return LoadObjInternal(attrib, shapes, materials, warn, err, sr,
                           &matFileReader, triangulate, default_vcols_fallback,
                           filename);
  }
#endif  // TINYOBJLOADER_USE_MMAP
}

bool LoadObj(attrib_t *attrib, std::vector<shape_t> *shapes,
             std::vector<material_t> *materials, std::string *warn,
             std::string *err, std::istream *inStream,
             MaterialReader *readMatFn /*= NULL*/, bool triangulate,
             bool default_vcols_fallback) {
  attrib->vertices.clear();
  attrib->vertex_weights.clear();
  attrib->normals.clear();
  attrib->texcoords.clear();
  attrib->texcoord_ws.clear();
  attrib->colors.clear();
  attrib->skin_weights.clear();
  shapes->clear();

  StreamReader sr(*inStream);
  return LoadObjInternal(attrib, shapes, materials, warn, err, sr,
                         readMatFn, triangulate, default_vcols_fallback);
}


static bool LoadObjWithCallbackInternal(StreamReader &sr,
                                        const callback_t &callback,
                                        void *user_data,
                                        MaterialReader *readMatFn,
                                        std::string *warn,
                                        std::string *err) {
  if (sr.has_errors()) {
    if (err) {
      (*err) += sr.get_errors();
    }
    return false;
  }

  // material
  std::set<std::string> material_filenames;
  std::map<std::string, int> material_map;
  int material_id = -1;

  std::vector<index_t> indices;
  std::vector<material_t> materials;
  std::vector<std::string> names;
  names.reserve(2);
  std::vector<const char *> names_out;

  // Handle BOM
  if (sr.remaining() >= 3 &&
      static_cast<unsigned char>(sr.peek()) == 0xEF &&
      static_cast<unsigned char>(sr.peek_at(1)) == 0xBB &&
      static_cast<unsigned char>(sr.peek_at(2)) == 0xBF) {
    sr.advance(3);
  }

  while (!sr.eof()) {
    sr.skip_space();
    if (sr.at_line_end()) { sr.skip_line(); continue; }
    if (sr.peek() == '#') { sr.skip_line(); continue; }

    // vertex
    if (sr.peek() == 'v' && (sr.peek_at(1) == ' ' || sr.peek_at(1) == '\t')) {
      sr.advance(2);
      real_t x, y, z;
      real_t r, g, b;

      int num_components = sr_parseVertexWithColor(&x, &y, &z, &r, &g, &b, sr, err, std::string());
      if (num_components < 0) {
        return false;
      }
      if (callback.vertex_cb) {
        callback.vertex_cb(user_data, x, y, z, r);
      }
      if (callback.vertex_color_cb) {
        bool found_color = (num_components == 6);
        callback.vertex_color_cb(user_data, x, y, z, r, g, b, found_color);
      }
      sr.skip_line();
      continue;
    }

    // normal
    if (sr.peek() == 'v' && sr.peek_at(1) == 'n' && (sr.peek_at(2) == ' ' || sr.peek_at(2) == '\t')) {
      sr.advance(3);
      real_t x, y, z;
      sr_parseReal3(&x, &y, &z, sr);
      if (callback.normal_cb) {
        callback.normal_cb(user_data, x, y, z);
      }
      sr.skip_line();
      continue;
    }

    // texcoord
    if (sr.peek() == 'v' && sr.peek_at(1) == 't' && (sr.peek_at(2) == ' ' || sr.peek_at(2) == '\t')) {
      sr.advance(3);
      real_t x, y, z;
      sr_parseReal3(&x, &y, &z, sr);
      if (callback.texcoord_cb) {
        callback.texcoord_cb(user_data, x, y, z);
      }
      sr.skip_line();
      continue;
    }

    // face
    if (sr.peek() == 'f' && (sr.peek_at(1) == ' ' || sr.peek_at(1) == '\t')) {
      sr.advance(2);
      sr.skip_space();

      indices.clear();
      size_t cf_loop_max = sr.remaining() + 1;
      size_t cf_loop_iter = 0;
      while (!sr.at_line_end() && sr.peek() != '#' &&
             cf_loop_iter < cf_loop_max) {
        vertex_index_t vi = sr_parseRawTriple(sr);

        index_t idx;
        idx.vertex_index = vi.v_idx;
        idx.normal_index = vi.vn_idx;
        idx.texcoord_index = vi.vt_idx;

        indices.push_back(idx);
        sr.skip_space_and_cr();
        cf_loop_iter++;
      }

      if (callback.index_cb && indices.size() > 0) {
        callback.index_cb(user_data, &indices.at(0),
                          static_cast<int>(indices.size()));
      }

      sr.skip_line();
      continue;
    }

    // use mtl
    if (sr.match("usemtl", 6) && (sr.peek_at(6) == ' ' || sr.peek_at(6) == '\t')) {
      sr.advance(6);
      std::string namebuf = sr_parseString(sr);

      int newMaterialId = -1;
      std::map<std::string, int>::const_iterator it =
          material_map.find(namebuf);
      if (it != material_map.end()) {
        newMaterialId = it->second;
      } else {
        if (warn && (!callback.usemtl_cb)) {
          (*warn) += "material [ " + namebuf + " ] not found in .mtl\n";
        }
      }

      if (newMaterialId != material_id) {
        material_id = newMaterialId;
      }

      if (callback.usemtl_cb) {
        callback.usemtl_cb(user_data, namebuf.c_str(), material_id);
      }

      sr.skip_line();
      continue;
    }

    // load mtl
    if (sr.match("mtllib", 6) && (sr.peek_at(6) == ' ' || sr.peek_at(6) == '\t')) {
      if (readMatFn) {
        sr.advance(7);

        std::string line_rest = trimTrailingWhitespace(sr.read_line());
        std::vector<std::string> filenames;
        SplitString(line_rest, ' ', '\\', filenames);
        RemoveEmptyTokens(&filenames);

        if (filenames.empty()) {
          if (warn) {
            (*warn) +=
                "Looks like empty filename for mtllib. Use default "
                "material. \n";
          }
        } else {
          bool found = false;
          for (size_t s = 0; s < filenames.size(); s++) {
            if (material_filenames.count(filenames[s]) > 0) {
              found = true;
              continue;
            }

            std::string warn_mtl;
            std::string err_mtl;
            bool ok = (*readMatFn)(filenames[s].c_str(), &materials,
                                   &material_map, &warn_mtl, &err_mtl);

            if (warn && (!warn_mtl.empty())) {
              (*warn) += warn_mtl;
            }

            if (err && (!err_mtl.empty())) {
              (*err) += err_mtl;
            }

            if (ok) {
              found = true;
              material_filenames.insert(filenames[s]);
              break;
            }
          }

          if (!found) {
            if (warn) {
              (*warn) +=
                  "Failed to load material file(s). Use default "
                  "material.\n";
            }
          } else {
            if (callback.mtllib_cb && !materials.empty()) {
              callback.mtllib_cb(user_data, &materials.at(0),
                                 static_cast<int>(materials.size()));
            }
          }
        }
      }

      sr.skip_line();
      continue;
    }

    // group name
    if (sr.peek() == 'g' && (sr.peek_at(1) == ' ' || sr.peek_at(1) == '\t')) {
      names.clear();

      size_t cg_loop_max = sr.remaining() + 1;
      size_t cg_loop_iter = 0;
      while (!sr.at_line_end() && sr.peek() != '#' &&
             cg_loop_iter < cg_loop_max) {
        std::string str = sr_parseString(sr);
        names.push_back(str);
        sr.skip_space_and_cr();
        cg_loop_iter++;
      }

      assert(names.size() > 0);

      if (callback.group_cb) {
        if (names.size() > 1) {
          names_out.resize(names.size() - 1);
          for (size_t j = 0; j < names_out.size(); j++) {
            names_out[j] = names[j + 1].c_str();
          }
          callback.group_cb(user_data, &names_out.at(0),
                            static_cast<int>(names_out.size()));

        } else {
          callback.group_cb(user_data, NULL, 0);
        }
      }

      sr.skip_line();
      continue;
    }

    // object name
    if (sr.peek() == 'o' && (sr.peek_at(1) == ' ' || sr.peek_at(1) == '\t')) {
      sr.advance(2);
      std::string object_name = sr.read_line();

      if (callback.object_cb) {
        callback.object_cb(user_data, object_name.c_str());
      }

      sr.skip_line();
      continue;
    }

#if 0  // @todo
    if (sr.peek() == 't' && (sr.peek_at(1) == ' ' || sr.peek_at(1) == '\t')) {
      tag_t tag;

      sr.advance(2);
      tag.name = sr_parseString(sr);

      tag_sizes ts = sr_parseTagTriple(sr);

      tag.intValues.resize(static_cast<size_t>(ts.num_ints));

      for (size_t i = 0; i < static_cast<size_t>(ts.num_ints); ++i) {
        tag.intValues[i] = sr_parseInt(sr);
      }

      tag.floatValues.resize(static_cast<size_t>(ts.num_reals));
      for (size_t i = 0; i < static_cast<size_t>(ts.num_reals); ++i) {
        tag.floatValues[i] = sr_parseReal(sr);
      }

      tag.stringValues.resize(static_cast<size_t>(ts.num_strings));
      for (size_t i = 0; i < static_cast<size_t>(ts.num_strings); ++i) {
        tag.stringValues[i] = sr_parseString(sr);
      }

      tags.push_back(tag);
    }
#endif

    // Ignore unknown command.
    sr.skip_line();
  }

  return true;
}

bool LoadObjWithCallback(std::istream &inStream, const callback_t &callback,
                         void *user_data /*= NULL*/,
                         MaterialReader *readMatFn /*= NULL*/,
                         std::string *warn, /* = NULL*/
                         std::string *err /*= NULL*/) {
  StreamReader sr(inStream);
  return LoadObjWithCallbackInternal(sr, callback, user_data, readMatFn,
                                     warn, err);
}

bool ObjReader::ParseFromFile(const std::string &filename,
                              const ObjReaderConfig &config) {
  std::string mtl_search_path;

  if (config.mtl_search_path.empty()) {
    //
    // split at last '/'(for unixish system) or '\\'(for windows) to get
    // the base directory of .obj file
    //
    size_t pos = filename.find_last_of("/\\");
    if (pos != std::string::npos) {
      mtl_search_path = filename.substr(0, pos);
    }
  } else {
    mtl_search_path = config.mtl_search_path;
  }

  valid_ = LoadObj(&attrib_, &shapes_, &materials_, &warning_, &error_,
                   filename.c_str(), mtl_search_path.c_str(),
                   config.triangulate, config.vertex_color);

  return valid_;
}

bool ObjReader::ParseFromString(const std::string &obj_text,
                                const std::string &mtl_text,
                                const ObjReaderConfig &config) {
  std::stringbuf obj_buf(obj_text);
  std::stringbuf mtl_buf(mtl_text);

  std::istream obj_ifs(&obj_buf);
  std::istream mtl_ifs(&mtl_buf);

  MaterialStreamReader mtl_ss(mtl_ifs);

  valid_ = LoadObj(&attrib_, &shapes_, &materials_, &warning_, &error_,
                   &obj_ifs, &mtl_ss, config.triangulate, config.vertex_color);

  return valid_;
}

// ===========================================================================
// Optimized API implementation (C++11+)
// ===========================================================================
// ---- ArenaAllocator implementation ----

void *ArenaAllocator::allocate(size_t bytes, size_t alignment) {
  if (bytes == 0) bytes = 1;

  // Try to allocate from current block
  if (head_) {
    size_t space = head_->capacity - head_->used;
    void *ptr = head_->data + head_->used;
    if (std::align(alignment, bytes, ptr, space)) {
      head_->used = static_cast<size_t>(static_cast<unsigned char *>(ptr) -
                                        head_->data) +
                    bytes;
      return ptr;
    }
  }

  // Guard against size_t overflow in bytes + alignment
  if (bytes > SIZE_MAX - alignment) {
#ifdef TINYOBJLOADER_ENABLE_EXCEPTION
    throw std::bad_alloc();
#else
    return nullptr;
#endif
  }

  // Need a new block
  Block *b = new_block(bytes + alignment);
  if (!b) return nullptr;
  size_t space = b->capacity;
  void *ptr = b->data;
  if (!std::align(alignment, bytes, ptr, space)) {
    // Defensive guard: the block was allocated with capacity >= bytes + alignment,
    // so alignment should always succeed.  This handles edge cases where the
    // capacity was insufficient due to unusual alignment requirements.
#ifdef TINYOBJLOADER_ENABLE_EXCEPTION
    throw std::bad_alloc();
#else
    return nullptr;
#endif
  }
  b->used =
      static_cast<size_t>(static_cast<unsigned char *>(ptr) - b->data) + bytes;
  return ptr;
}

void ArenaAllocator::reset() { destroy(); }

ArenaAllocator::Block *ArenaAllocator::new_block(size_t min_bytes) {
  size_t cap = (min_bytes > default_block_size_) ? min_bytes
                                                 : default_block_size_;
#ifdef TINYOBJLOADER_ENABLE_EXCEPTION
  // Allocate data buffer first: if this throws std::bad_alloc, no Block
  // struct is leaked.  If the subsequent Block allocation throws (very
  // unlikely for a small POD), the data buffer is cleaned up.
  unsigned char *data = new unsigned char[cap];
  Block *b;
  try {
    b = new Block;
  } catch (...) {
    delete[] data;
    throw;
  }
  b->data = data;
#else
  Block *b = new (std::nothrow) Block;
  if (!b) return nullptr;
  b->data = new (std::nothrow) unsigned char[cap];
  if (!b->data) { delete b; return nullptr; }
#endif
  b->capacity = cap;
  b->used = 0;
  b->next = head_;
  head_ = b;
  return b;
}

void ArenaAllocator::destroy() {
  Block *b = head_;
  while (b) {
    Block *next = b->next;
    delete[] b->data;
    delete b;
    b = next;
  }
  head_ = nullptr;
}

// ---- Optimized parser internals ----

namespace opt_internal {

static const int kOptMaxThreads = 32;

struct LineInfo {
  size_t pos;
  size_t len;
};

#define TINYOBJ_OPT_IS_SPACE(x) (((x) == ' ') || ((x) == '\t'))
#define TINYOBJ_OPT_IS_DIGIT(x) \
  (static_cast<unsigned int>((x) - '0') < static_cast<unsigned int>(10))
#define TINYOBJ_OPT_IS_NEW_LINE(x) \
  (((x) == '\r') || ((x) == '\n') || ((x) == '\0'))

static inline void opt_skip_space(const char **token) {
  while ((**token) == ' ' || (**token) == '\t') {
    (*token)++;
  }
}

static inline int opt_until_space(const char *token) {
  const char *p = token;
  while (p[0] != '\0' && p[0] != ' ' && p[0] != '\t' && p[0] != '\r' &&
         p[0] != '\n') {
    p++;
  }
  return static_cast<int>(p - token);
}

static inline int opt_my_atoi(const char *c) {
  unsigned int value = 0;
  int sign = 1;
  if (*c == '+' || *c == '-') {
    if (*c == '-') sign = -1;
    c++;
  }
  while ((*c >= '0') && (*c <= '9')) {
    const unsigned int digit = static_cast<unsigned int>(*c - '0');
    const unsigned int limit = (sign < 0)
                                   ? static_cast<unsigned int>(INT_MAX) + 1u
                                   : static_cast<unsigned int>(INT_MAX);
    if (value > (limit / 10u) ||
        (value == (limit / 10u) && digit > (limit % 10u))) {
      return (sign < 0) ? INT_MIN : INT_MAX;
    }
    value = value * 10u + digit;
    c++;
  }
  if (sign < 0) {
    if (value == static_cast<unsigned int>(INT_MAX) + 1u) {
      return INT_MIN;
    }
    return -static_cast<int>(value);
  }
  return static_cast<int>(value);
}

static inline const char *opt_find_index_token_end(const char *token) {
  const char *end = token;
  while (*end != '\0' && *end != '/' && *end != ' ' && *end != '\t' &&
         *end != '\r' && *end != '\n') {
    end++;
  }
  return end;
}

static inline bool opt_tryParseIndexToken(const char *token, const char *end,
                                          int *value) {
  if (!value || !token || !end || token >= end) return false;

  const char *cursor = token;
  if (*cursor == '+' || *cursor == '-') {
    cursor++;
  }
  if (cursor >= end) return false;
  while (cursor < end) {
    if (!TINYOBJ_OPT_IS_DIGIT(*cursor)) return false;
    cursor++;
  }

  *value = opt_my_atoi(token);
  return true;
}

static inline bool opt_resolveIndexLikeLegacy(int idx, int n, int *ret,
                                              bool allow_zero) {
  if (!ret) return false;
  if (idx > 0) {
    (*ret) = idx - 1;
    return true;
  }
  if (idx == 0) {
    (*ret) = -1;
    return allow_zero;
  }

  (*ret) = n + idx;
  return ((*ret) >= 0);
}

static inline void opt_appendZeroIndexWarning(std::string *warn,
                                              const std::string &source_name,
                                              size_t line_num) {
  if (!warn) return;

  std::stringstream ss;
  ss << source_name << ":" << line_num
     << ": warning: zero value index found (will have a value of -1 for "
        "normal and tex indices)\n";
  (*warn) += ss.str();
}

static inline bool opt_validateAndResolveFaceIndexLikeLegacy(
    int raw_idx, int n, bool allow_zero, const std::string &source_name,
    size_t line_num, std::string *warn, int *resolved_idx) {
  if (raw_idx > 0) {
    if (resolved_idx) {
      (*resolved_idx) = raw_idx - 1;
    }
    return true;
  }

  if (raw_idx == 0) {
    opt_appendZeroIndexWarning(warn, source_name, line_num);
    if (resolved_idx) {
      (*resolved_idx) = -1;
    }
    return allow_zero;
  }

  if (resolved_idx) {
    (*resolved_idx) = n + raw_idx;
    return ((*resolved_idx) >= 0);
  }

  return ((n + raw_idx) >= 0);
}

static inline void opt_updateGreatestIndex(int idx, int *greatest) {
  if (!greatest) return;
  if (idx > *greatest) {
    *greatest = idx;
  }
}

static inline void opt_appendOutOfBoundsWarnings(std::string *warn,
                                                 int greatest_v_idx,
                                                 int greatest_vn_idx,
                                                 int greatest_vt_idx,
                                                 int num_vertices,
                                                 int num_normals,
                                                 int num_texcoords,
                                                 size_t line_num) {
  if (!warn) return;

  if (greatest_v_idx >= num_vertices) {
    std::stringstream ss;
    ss << "Vertex indices out of bounds (line " << line_num << ".)\n\n";
    (*warn) += ss.str();
  }
  if (greatest_vn_idx >= num_normals) {
    std::stringstream ss;
    ss << "Vertex normal indices out of bounds (line " << line_num << ".)\n\n";
    (*warn) += ss.str();
  }
  if (greatest_vt_idx >= num_texcoords) {
    std::stringstream ss;
    ss << "Vertex texcoord indices out of bounds (line " << line_num
       << ".)\n\n";
    (*warn) += ss.str();
  }
}

// Hand-written fallback double parser.  Compiled only when fast_float is
// disabled (TINYOBJLOADER_DISABLE_FAST_FLOAT); otherwise callers use
// fast_float::from_chars directly and this function is not needed.
#ifdef TINYOBJLOADER_DISABLE_FAST_FLOAT
static bool opt_tryParseDouble(const char *s, const char *s_end,
                               double *result) {
  if (s >= s_end) return false;

  // Handle nan/inf keywords with OBJ-compatible replacement values.
  {
    const char *p = s;
    bool neg = false;
    if (p < s_end && *p == '-') { neg = true; ++p; }
    else if (p < s_end && *p == '+') { ++p; }
    if (p < s_end) {
      char fc = *p;
      if (fc >= 'A' && fc <= 'Z') fc += 32;
      if (fc == 'n' && (p + 2 < s_end)) {
        char c1 = p[1], c2 = p[2];
        if (c1 >= 'A' && c1 <= 'Z') c1 += 32;
        if (c2 >= 'A' && c2 <= 'Z') c2 += 32;
        if (c1 == 'a' && c2 == 'n') {
          *result = 0.0;
          return true;
        }
      }
      if (fc == 'i' && (p + 2 < s_end)) {
        char c1 = p[1], c2 = p[2];
        if (c1 >= 'A' && c1 <= 'Z') c1 += 32;
        if (c2 >= 'A' && c2 <= 'Z') c2 += 32;
        if (c1 == 'n' && c2 == 'f') {
          *result = neg ? std::numeric_limits<double>::lowest()
                        : (std::numeric_limits<double>::max)();
          return true;
        }
      }
    }
  }

  double mantissa = 0.0;
  int exponent = 0;
  char sign = '+';
  char exp_sign = '+';
  const char *curr = s;
  int read = 0;
  bool end_not_reached = false;
  bool has_leading_decimal = false;

  if (*curr == '+' || *curr == '-') {
    sign = *curr;
    curr++;
  }

  if (curr == s_end) return false;

  if (*curr == '.') {
    has_leading_decimal = true;
  } else if (!TINYOBJ_OPT_IS_DIGIT(*curr)) {
    return false;
  }

  end_not_reached = (curr != s_end);
  if (!has_leading_decimal) {
    while (end_not_reached && TINYOBJ_OPT_IS_DIGIT(*curr)) {
      mantissa *= 10;
      mantissa += static_cast<int>(*curr - '0');
      curr++;
      read++;
      end_not_reached = (curr != s_end);
    }
    if (read == 0) return false;
  }
  if (!end_not_reached) goto opt_assemble;

  if (*curr == '.') {
    curr++;
    end_not_reached = (curr != s_end);
    double frac_scale = 0.1;
    while (end_not_reached && TINYOBJ_OPT_IS_DIGIT(*curr)) {
      mantissa += static_cast<int>(*curr - '0') * frac_scale;
      frac_scale *= 0.1;
      read++;
      curr++;
      end_not_reached = (curr != s_end);
    }
    if (has_leading_decimal && read == 0) return false;
  } else if (*curr != 'e' && *curr != 'E') {
    goto opt_assemble;
  }

  if (!end_not_reached) goto opt_assemble;

  if (*curr == 'e' || *curr == 'E') {
    curr++;
    end_not_reached = (curr != s_end);
    if (!end_not_reached) return false;
    if (*curr == '+' || *curr == '-') {
      exp_sign = *curr;
      curr++;
      end_not_reached = (curr != s_end);
    } else if (!TINYOBJ_OPT_IS_DIGIT(*curr)) {
      return false;
    }
    read = 0;
    end_not_reached = (curr != s_end);
    while (end_not_reached && TINYOBJ_OPT_IS_DIGIT(*curr)) {
      // Clamp to avoid signed integer overflow (UB).  |exponent| > 308
      // already exceeds double range, so further digits are irrelevant.
      if (exponent < 0x7FFFFFF) {
        exponent *= 10;
        exponent += static_cast<int>(*curr - '0');
      }
      curr++;
      read++;
      end_not_reached = (curr != s_end);
    }
    exponent *= (exp_sign == '+' ? 1 : -1);
    if (read == 0) return false;
  }

opt_assemble:
  *result = (sign == '+' ? 1.0 : -1.0) *
            (exponent ? std::ldexp(mantissa * std::pow(5.0, exponent), exponent)
                      : mantissa);
  return true;
}
#endif  // TINYOBJLOADER_DISABLE_FAST_FLOAT

struct opt_index_t {
  int vertex_index, texcoord_index, normal_index;
  // Sentinel for "field not present" to distinguish from OBJ relative index -1.
  // Using expression form for C++11 static const initializer compatibility.
  static const int kNotPresent = -2147483647 - 1;  // == std::numeric_limits<int>::min()
  opt_index_t()
      : vertex_index(kNotPresent),
        texcoord_index(kNotPresent),
        normal_index(kNotPresent) {}
  opt_index_t(int vi, int ti, int ni)
      : vertex_index(vi), texcoord_index(ti), normal_index(ni) {}
};

static bool opt_parseRawTriple(const char **token, opt_index_t *ret) {
  if (!token || !ret) return false;

  opt_index_t vi;
  const char *segment_end = opt_find_index_token_end(*token);
  if (!opt_tryParseIndexToken(*token, segment_end, &vi.vertex_index)) {
    return false;
  }
  *token = segment_end;
  if (**token != '/') {
    *ret = vi;
    return true;
  }
  (*token)++;

  if (**token == '/') {
    (*token)++;
    segment_end = opt_find_index_token_end(*token);
    if (!opt_tryParseIndexToken(*token, segment_end, &vi.normal_index)) {
      return false;
    }
    *token = segment_end;
    *ret = vi;
    return true;
  }

  segment_end = opt_find_index_token_end(*token);
  if (!opt_tryParseIndexToken(*token, segment_end, &vi.texcoord_index)) {
    return false;
  }
  *token = segment_end;
  if (**token != '/') {
    *ret = vi;
    return true;
  }
  (*token)++;
  segment_end = opt_find_index_token_end(*token);
  if (!opt_tryParseIndexToken(*token, segment_end, &vi.normal_index)) {
    return false;
  }
  *token = segment_end;
  *ret = vi;
  return true;
}

static inline int opt_length_until_newline(const char *token, size_t n) {
  size_t len = 0;
  for (len = 0; len < n; len++) {
    if (token[len] == '\n' || token[len] == '\r') break;
  }
  // Trim trailing whitespace
  while (len > 0 && (token[len - 1] == ' ' || token[len - 1] == '\t')) {
    len--;
  }
  return static_cast<int>(len);
}

static inline int opt_length_until_token_or_comment(const char *token, size_t n) {
  size_t len = 0;
  for (len = 0; len < n; len++) {
    const char c = token[len];
    if (c == '\n' || c == '\r' || c == ' ' || c == '\t') break;
  }
  return static_cast<int>(len);
}

static inline std::string opt_parseGroupName(const char *token, size_t n) {
  std::string name;
  size_t i = 0;
  while (i < n) {
    while (i < n && (token[i] == ' ' || token[i] == '\t')) {
      i++;
    }
    if (i >= n || token[i] == '\n' || token[i] == '\r' || token[i] == '\0' ||
        token[i] == '#') {
      break;
    }

    const size_t start = i;
    while (i < n && token[i] != '\n' && token[i] != '\r' &&
           token[i] != '\0' && token[i] != ' ' && token[i] != '\t') {
      i++;
    }

    if (!name.empty()) {
      name.push_back(' ');
    }
    name.append(token + start, i - start);
  }
  return name;
}

static inline bool opt_tryParseFloatToken(real_t *out, const char **token) {
  if (!out || !token) return false;
  const char *cursor = *token;
  opt_skip_space(&cursor);
  if (TINYOBJ_OPT_IS_NEW_LINE(cursor[0]) || cursor[0] == '#' || cursor[0] == '\0') {
    return false;
  }
  const char *end = cursor;
  while (!TINYOBJ_OPT_IS_NEW_LINE(end[0]) && end[0] != '#' && end[0] != ' ' &&
         end[0] != '\t' && end[0] != '\0') {
    end++;
  }
#ifndef TINYOBJLOADER_DISABLE_FAST_FLOAT
  // Handle nan/inf with OBJ-compatible values before fast_float.
  {
    const char *q = cursor;
    if (q < end && (*q == '+' || *q == '-')) ++q;
    if (q < end) {
      char fc = *q;
      if (fc >= 'A' && fc <= 'Z') fc += 32;
      if (fc == 'n' || fc == 'i') {
        double special_val;
        const char *end_ptr;
        if (detail_fp::tryParseNanInf(cursor, end, &special_val, &end_ptr)) {
          *out = static_cast<real_t>(special_val);
          *token = end;
          return true;
        }
      }
    }
  }
  // Parse directly to real_t (float or double) via fast_float — avoids
  // the double→float conversion and is ~3-4x faster than the hand-rolled parser.
  real_t tmp;
  auto r = fast_float::from_chars(cursor, end, tmp,
      fast_float::chars_format::general |
      fast_float::chars_format::allow_leading_plus);
  if (r.ec == tinyobj_ff::ff_errc::ok) {
    *out = tmp;
    *token = end;
    return true;
  }
  return false;
#else
  double val = 0.0;
  if (!opt_tryParseDouble(cursor, end, &val)) {
    return false;
  }
  *out = static_cast<real_t>(val);
  *token = end;
  return true;
#endif
}

static inline int opt_count_remaining_scalars(const char *token) {
  int count = 0;
  const char *cursor = token;
  while (true) {
    opt_skip_space(&cursor);
    if (TINYOBJ_OPT_IS_NEW_LINE(cursor[0]) || cursor[0] == '#' ||
        cursor[0] == '\0') {
      break;
    }
    count++;
    while (!TINYOBJ_OPT_IS_NEW_LINE(cursor[0]) && cursor[0] != '#' &&
           cursor[0] != ' ' && cursor[0] != '\t' && cursor[0] != '\0') {
      cursor++;
    }
  }
  return count;
}

static inline bool opt_is_comment_start(const char *token) {
  return token[0] == '#';
}

enum OptCommandType {
  OPT_CMD_EMPTY,
  OPT_CMD_V,
  OPT_CMD_VN,
  OPT_CMD_VT,
  OPT_CMD_F,
  OPT_CMD_G,
  OPT_CMD_O,
  OPT_CMD_USEMTL,
  OPT_CMD_MTLLIB,
  OPT_CMD_S
};

struct OptCommand {
  static const unsigned int kInlineIndexCapacity = 24;

  real_t vx, vy, vz, vw;
  real_t vc_r, vc_g, vc_b;
  real_t nx, ny, nz;
  real_t tx, ty, tw;
  bool has_vertex_weight;
  bool has_vertex_color;
  bool has_texcoord_w;

  opt_index_t f_inline[kInlineIndexCapacity];
  unsigned int f_count;
  unsigned int face_vertex_count;
  unsigned int emitted_face_count;
  unsigned int emitted_face_verts;
  std::vector<opt_index_t> f_heap;

  const char *group_name;
  unsigned int group_name_len;
  std::string group_name_storage;
  const char *object_name;
  unsigned int object_name_len;
  const char *material_name;
  unsigned int material_name_len;
  const char *mtllib_name;
  unsigned int mtllib_name_len;
  size_t source_line;
  bool group_name_empty;
  bool degenerate_face;
  int resolved_material_id;
  unsigned int smoothing_group_id;

  OptCommandType type;

  OptCommand()
      : vx(0), vy(0), vz(0), vw(1),
        vc_r(1), vc_g(1), vc_b(1),
        nx(0), ny(0), nz(0),
        tx(0), ty(0), tw(0),
        has_vertex_weight(false), has_vertex_color(false),
        has_texcoord_w(false),
        f_count(0), face_vertex_count(0), emitted_face_count(0),
        emitted_face_verts(0),
        group_name(nullptr), group_name_len(0),
        object_name(nullptr), object_name_len(0),
        material_name(nullptr), material_name_len(0),
        mtllib_name(nullptr), mtllib_name_len(0),
        source_line(0), group_name_empty(false),
        degenerate_face(false),
        resolved_material_id(-1),
        smoothing_group_id(0),
        type(OPT_CMD_EMPTY) {}

  const opt_index_t *face_indices() const {
    return f_heap.empty() ? f_inline : f_heap.data();
  }
};

struct OptCommandCount {
  size_t num_v, num_vn, num_vt, num_f, num_indices;
  OptCommandCount() : num_v(0), num_vn(0), num_vt(0), num_f(0), num_indices(0) {}
};

// Compact face command — replaces OptCommand for 'f' lines
struct OptFaceCmd {
  static const unsigned int kInlineCap = 4;
  opt_index_t f_inline[kInlineCap];  // 48 bytes (covers tri + quad)
  std::vector<opt_index_t> f_heap;   // overflow for >4 vertex faces
  uint32_t face_vertex_count;
  uint32_t emitted_face_count;
  uint32_t emitted_face_verts;
  uint32_t f_count;
  uint32_t source_line;
  uint32_t v_count_before;   // running v count when face was parsed (within thread)
  uint32_t vn_count_before;
  uint32_t vt_count_before;
  bool degenerate;

  const opt_index_t *face_indices() const {
    return f_heap.empty() ? f_inline : f_heap.data();
  }

  OptFaceCmd()
      : face_vertex_count(0), emitted_face_count(0), emitted_face_verts(0),
        f_count(0), source_line(0), v_count_before(0), vn_count_before(0),
        vt_count_before(0), degenerate(false) {}
};

// Meta command — for g, o, usemtl, mtllib, s lines
struct OptMetaCmd {
  OptCommandType type;
  uint32_t source_line;
  const char *str_ptr;
  uint32_t str_len;
  std::string str_storage;
  bool group_name_empty;
  int resolved_material_id;
  uint32_t smoothing_group_id;

  OptMetaCmd()
      : type(OPT_CMD_EMPTY), source_line(0), str_ptr(NULL), str_len(0),
        group_name_empty(false), resolved_material_id(-1),
        smoothing_group_id(0) {}
};

// Sequence entry — ordering of faces and meta commands within a thread
struct OptSeqEntry {
  enum Kind : unsigned char { SEQ_FACE = 0, SEQ_META = 1 };
  Kind kind;
  uint32_t index;
};

// Per-thread parsed data — replaces vector<OptCommand>
struct OptThreadData {
  // Vertex positions: 3 floats per vertex, contiguous
  std::vector<real_t> v_pos;
  // Vertex weights: 1 per vertex, lazily allocated (only if any has weight)
  std::vector<real_t> v_weight;
  // Vertex colors: 3 per vertex, lazily allocated
  std::vector<real_t> v_color;

  // Normal data: 3 floats per normal
  std::vector<real_t> vn_data;

  // Texcoord data: 2 floats per texcoord
  std::vector<real_t> vt_data;
  // Texcoord w: 1 per texcoord, lazily allocated
  std::vector<real_t> vt_w;

  // Non-vertex commands
  std::vector<OptFaceCmd> faces;
  std::vector<OptMetaCmd> metas;
  std::vector<OptSeqEntry> seq;

  // Counts
  size_t num_v, num_vn, num_vt;
  size_t num_f_indices;   // total index_t entries for faces
  size_t num_f_faces;     // total emitted faces

  // Flags
  bool saw_any_color, saw_missing_color;
  bool saw_any_weight;
  bool saw_any_texcoord_w;
  bool saw_any_smoothing;

  // Error tracking
  size_t error_line;
  std::string error_message;

  OptThreadData()
      : num_v(0), num_vn(0), num_vt(0), num_f_indices(0), num_f_faces(0),
        saw_any_color(false), saw_missing_color(false),
        saw_any_weight(false), saw_any_texcoord_w(false),
        saw_any_smoothing(false), error_line(0) {}
};

static inline bool opt_is_valid_face_vertex(const std::vector<real_t> &vertices,
                                            const index_t &idx) {
  if (idx.vertex_index < 0) return false;
  const size_t vi = static_cast<size_t>(idx.vertex_index);
  return ((3 * vi + 2) < vertices.size());
}

static inline size_t opt_triangulate_face(const std::vector<real_t> &vertices,
                                          const index_t *face,
                                          size_t face_count,
                                          index_t *dst) {
  if (face_count < 3) return 0;
  if (face_count == 3) {
    dst[0] = face[0];
    dst[1] = face[1];
    dst[2] = face[2];
    return 3;
  }

  for (size_t i = 0; i < face_count; i++) {
    if (!opt_is_valid_face_vertex(vertices, face[i])) {
      return 0;
    }
  }

  if (face_count == 4) {
    const size_t vi0 = static_cast<size_t>(face[0].vertex_index);
    const size_t vi1 = static_cast<size_t>(face[1].vertex_index);
    const size_t vi2 = static_cast<size_t>(face[2].vertex_index);
    const size_t vi3 = static_cast<size_t>(face[3].vertex_index);

    const real_t v0x = vertices[vi0 * 3 + 0];
    const real_t v0y = vertices[vi0 * 3 + 1];
    const real_t v0z = vertices[vi0 * 3 + 2];
    const real_t v1x = vertices[vi1 * 3 + 0];
    const real_t v1y = vertices[vi1 * 3 + 1];
    const real_t v1z = vertices[vi1 * 3 + 2];
    const real_t v2x = vertices[vi2 * 3 + 0];
    const real_t v2y = vertices[vi2 * 3 + 1];
    const real_t v2z = vertices[vi2 * 3 + 2];
    const real_t v3x = vertices[vi3 * 3 + 0];
    const real_t v3y = vertices[vi3 * 3 + 1];
    const real_t v3z = vertices[vi3 * 3 + 2];

    const real_t e02x = v2x - v0x;
    const real_t e02y = v2y - v0y;
    const real_t e02z = v2z - v0z;
    const real_t e13x = v3x - v1x;
    const real_t e13y = v3y - v1y;
    const real_t e13z = v3z - v1z;
    const real_t sqr02 = e02x * e02x + e02y * e02y + e02z * e02z;
    const real_t sqr13 = e13x * e13x + e13y * e13y + e13z * e13z;

    if (sqr02 < sqr13) {
      dst[0] = face[0];
      dst[1] = face[1];
      dst[2] = face[2];
      dst[3] = face[0];
      dst[4] = face[2];
      dst[5] = face[3];
    } else {
      dst[0] = face[0];
      dst[1] = face[1];
      dst[2] = face[3];
      dst[3] = face[1];
      dst[4] = face[2];
      dst[5] = face[3];
    }
    return 6;
  }

  std::vector<index_t> remaining(face, face + face_count);
  size_t axes[2] = {1, 2};
  for (size_t k = 0; k < face_count; ++k) {
    const size_t vi0 = static_cast<size_t>(face[(k + 0) % face_count].vertex_index);
    const size_t vi1 = static_cast<size_t>(face[(k + 1) % face_count].vertex_index);
    const size_t vi2 = static_cast<size_t>(face[(k + 2) % face_count].vertex_index);
    const real_t v0x = vertices[vi0 * 3 + 0];
    const real_t v0y = vertices[vi0 * 3 + 1];
    const real_t v0z = vertices[vi0 * 3 + 2];
    const real_t v1x = vertices[vi1 * 3 + 0];
    const real_t v1y = vertices[vi1 * 3 + 1];
    const real_t v1z = vertices[vi1 * 3 + 2];
    const real_t v2x = vertices[vi2 * 3 + 0];
    const real_t v2y = vertices[vi2 * 3 + 1];
    const real_t v2z = vertices[vi2 * 3 + 2];
    const real_t e0x = v1x - v0x;
    const real_t e0y = v1y - v0y;
    const real_t e0z = v1z - v0z;
    const real_t e1x = v2x - v1x;
    const real_t e1y = v2y - v1y;
    const real_t e1z = v2z - v1z;
    const real_t cx = std::fabs(e0y * e1z - e0z * e1y);
    const real_t cy = std::fabs(e0z * e1x - e0x * e1z);
    const real_t cz = std::fabs(e0x * e1y - e0y * e1x);
    const real_t epsilon = std::numeric_limits<real_t>::epsilon();
    if (cx > epsilon || cy > epsilon || cz > epsilon) {
      if (!(cx > cy && cx > cz)) {
        axes[0] = 0;
        if (cz > cx && cz > cy) {
          axes[1] = 1;
        }
      }
      break;
    }
  }

  size_t out = 0;
  size_t guess_vert = 0;
  size_t remaining_iterations = remaining.size();
  size_t previous_remaining_vertices = remaining.size();
  while (remaining.size() > 3 && remaining_iterations > 0) {
    const size_t npolys = remaining.size();
    if (guess_vert >= npolys) {
      guess_vert -= npolys;
    }

    if (previous_remaining_vertices != npolys) {
      previous_remaining_vertices = npolys;
      remaining_iterations = npolys;
    } else {
      remaining_iterations--;
    }

    index_t ind[3];
    real_t vx[3];
    real_t vy[3];
    for (size_t k = 0; k < 3; k++) {
      ind[k] = remaining[(guess_vert + k) % npolys];
      const size_t vi = static_cast<size_t>(ind[k].vertex_index);
      vx[k] = vertices[vi * 3 + axes[0]];
      vy[k] = vertices[vi * 3 + axes[1]];
    }

    const real_t e0x = vx[1] - vx[0];
    const real_t e0y = vy[1] - vy[0];
    const real_t e1x = vx[2] - vx[1];
    const real_t e1y = vy[2] - vy[1];
    const real_t cross_val = e0x * e1y - e0y * e1x;
    const real_t area =
        (vx[0] * vy[1] - vy[0] * vx[1]) * static_cast<real_t>(0.5);
    if (cross_val * area < static_cast<real_t>(0.0)) {
      guess_vert += 1;
      continue;
    }

    bool overlap = false;
    for (size_t other_vert = 3; other_vert < npolys; ++other_vert) {
      const size_t idx = (guess_vert + other_vert) % npolys;
      const size_t ovi = static_cast<size_t>(remaining[idx].vertex_index);
      const real_t tx = vertices[ovi * 3 + axes[0]];
      const real_t ty = vertices[ovi * 3 + axes[1]];
      if (pnpoly(3, vx, vy, tx, ty)) {
        overlap = true;
        break;
      }
    }

    if (overlap) {
      guess_vert += 1;
      continue;
    }

    dst[out++] = ind[0];
    dst[out++] = ind[1];
    dst[out++] = ind[2];
    remaining.erase(remaining.begin() +
                    static_cast<std::ptrdiff_t>((guess_vert + 1) % npolys));
  }

  if (remaining.size() == 3) {
    dst[out++] = remaining[0];
    dst[out++] = remaining[1];
    dst[out++] = remaining[2];
  }

  return out;
}

// Pointer+size overloads for TypedArray-based API
static inline bool opt_is_valid_face_vertex(const real_t * /*vertices*/,
                                            size_t vertices_size,
                                            const index_t &idx) {
  if (idx.vertex_index < 0) return false;
  const size_t vi = static_cast<size_t>(idx.vertex_index);
  return ((3 * vi + 2) < vertices_size);
}

static inline size_t opt_triangulate_face(const real_t *vertices,
                                          size_t vertices_size,
                                          const index_t *face,
                                          size_t face_count,
                                          index_t *dst) {
  if (face_count < 3) return 0;
  if (face_count == 3) {
    dst[0] = face[0]; dst[1] = face[1]; dst[2] = face[2];
    return 3;
  }

  for (size_t i = 0; i < face_count; i++) {
    if (!opt_is_valid_face_vertex(vertices, vertices_size, face[i])) return 0;
  }

  if (face_count == 4) {
    const size_t vi0 = static_cast<size_t>(face[0].vertex_index);
    const size_t vi1 = static_cast<size_t>(face[1].vertex_index);
    const size_t vi2 = static_cast<size_t>(face[2].vertex_index);
    const size_t vi3 = static_cast<size_t>(face[3].vertex_index);
    const real_t e02x = vertices[vi2*3+0] - vertices[vi0*3+0];
    const real_t e02y = vertices[vi2*3+1] - vertices[vi0*3+1];
    const real_t e02z = vertices[vi2*3+2] - vertices[vi0*3+2];
    const real_t e13x = vertices[vi3*3+0] - vertices[vi1*3+0];
    const real_t e13y = vertices[vi3*3+1] - vertices[vi1*3+1];
    const real_t e13z = vertices[vi3*3+2] - vertices[vi1*3+2];
    const real_t sqr02 = e02x*e02x + e02y*e02y + e02z*e02z;
    const real_t sqr13 = e13x*e13x + e13y*e13y + e13z*e13z;
    if (sqr02 < sqr13) {
      dst[0]=face[0]; dst[1]=face[1]; dst[2]=face[2];
      dst[3]=face[0]; dst[4]=face[2]; dst[5]=face[3];
    } else {
      dst[0]=face[0]; dst[1]=face[1]; dst[2]=face[3];
      dst[3]=face[1]; dst[4]=face[2]; dst[5]=face[3];
    }
    return 6;
  }

  // General polygon — ear clipping
  std::vector<index_t> remaining(face, face + face_count);
  size_t axes[2] = {1, 2};
  for (size_t k = 0; k < face_count; ++k) {
    const size_t vi0 = static_cast<size_t>(face[(k+0)%face_count].vertex_index);
    const size_t vi1 = static_cast<size_t>(face[(k+1)%face_count].vertex_index);
    const size_t vi2 = static_cast<size_t>(face[(k+2)%face_count].vertex_index);
    const real_t e0x = vertices[vi1*3+0]-vertices[vi0*3+0];
    const real_t e0y = vertices[vi1*3+1]-vertices[vi0*3+1];
    const real_t e0z = vertices[vi1*3+2]-vertices[vi0*3+2];
    const real_t e1x = vertices[vi2*3+0]-vertices[vi1*3+0];
    const real_t e1y = vertices[vi2*3+1]-vertices[vi1*3+1];
    const real_t e1z = vertices[vi2*3+2]-vertices[vi1*3+2];
    const real_t cx = std::fabs(e0y*e1z - e0z*e1y);
    const real_t cy = std::fabs(e0z*e1x - e0x*e1z);
    const real_t cz = std::fabs(e0x*e1y - e0y*e1x);
    const real_t epsilon = std::numeric_limits<real_t>::epsilon();
    if (cx > epsilon || cy > epsilon || cz > epsilon) {
      if (!(cx > cy && cx > cz)) {
        axes[0] = 0;
        if (cz > cx && cz > cy) axes[1] = 1;
      }
      break;
    }
  }

  size_t out = 0;
  size_t guess_vert = 0;
  size_t remaining_iterations = remaining.size();
  size_t previous_remaining_vertices = remaining.size();
  while (remaining.size() > 3 && remaining_iterations > 0) {
    const size_t npolys = remaining.size();
    if (guess_vert >= npolys) guess_vert -= npolys;
    if (previous_remaining_vertices != npolys) {
      previous_remaining_vertices = npolys;
      remaining_iterations = npolys;
    } else {
      remaining_iterations--;
    }
    index_t ind[3];
    real_t vx[3], vy[3];
    for (size_t k = 0; k < 3; k++) {
      ind[k] = remaining[(guess_vert + k) % npolys];
      const size_t vi = static_cast<size_t>(ind[k].vertex_index);
      vx[k] = vertices[vi*3+axes[0]];
      vy[k] = vertices[vi*3+axes[1]];
    }
    const real_t e0x_ = vx[1]-vx[0], e0y_ = vy[1]-vy[0];
    const real_t e1x_ = vx[2]-vx[1], e1y_ = vy[2]-vy[1];
    const real_t cross_val = e0x_*e1y_ - e0y_*e1x_;
    const real_t area = (vx[0]*vy[1] - vy[0]*vx[1]) * static_cast<real_t>(0.5);
    if (cross_val * area < static_cast<real_t>(0.0)) {
      guess_vert += 1; continue;
    }
    bool overlap = false;
    for (size_t other_vert = 3; other_vert < npolys; ++other_vert) {
      const size_t idx = (guess_vert + other_vert) % npolys;
      const size_t ovi = static_cast<size_t>(remaining[idx].vertex_index);
      const real_t ptx = vertices[ovi*3+axes[0]];
      const real_t pty = vertices[ovi*3+axes[1]];
      if (pnpoly(3, vx, vy, ptx, pty)) { overlap = true; break; }
    }
    if (overlap) { guess_vert += 1; continue; }
    dst[out++] = ind[0]; dst[out++] = ind[1]; dst[out++] = ind[2];
    remaining.erase(remaining.begin() +
                    static_cast<std::ptrdiff_t>((guess_vert + 1) % npolys));
  }
  if (remaining.size() == 3) {
    dst[out++] = remaining[0]; dst[out++] = remaining[1]; dst[out++] = remaining[2];
  }
  return out;
}

static bool opt_parseLine(OptCommand *command, const char *p, size_t p_len,
                          bool triangulate, bool *parse_error,
                          std::string *parse_error_message) {
  // Parse directly from the original buffer without copying.
  // The caller guarantees that p[p_len] is '\n' (or a sentinel),
  // so character-scanning helpers that stop on '\n' are safe.
  if (parse_error) *parse_error = false;
  if (parse_error_message) parse_error_message->clear();
  const char *token = p;
  command->type = OPT_CMD_EMPTY;
  opt_skip_space(&token);

  if (TINYOBJ_OPT_IS_NEW_LINE(token[0]) || token[0] == '#') return false;

  // vertex
  if (token[0] == 'v' && TINYOBJ_OPT_IS_SPACE(token[1])) {
    token += 2;
    real_t x = 0, y = 0, z = 0;
    real_t r = real_t(1.0), g = real_t(1.0), b = real_t(1.0);
    real_t w = real_t(1.0);
    if (!opt_tryParseFloatToken(&x, &token) ||
        !opt_tryParseFloatToken(&y, &token) ||
        !opt_tryParseFloatToken(&z, &token)) {
      if (parse_error) *parse_error = true;
      if (parse_error_message) *parse_error_message = "failed to parse `v' line";
      return false;
    }
    const char *extra_token = token;
    opt_skip_space(&extra_token);
    const int extra_components = opt_count_remaining_scalars(extra_token);
    if (extra_components == 1) {
      // v x y z w  (4 total) — weight only
      const char *weight_cursor = extra_token;
      if (opt_tryParseFloatToken(&w, &weight_cursor)) {
        command->has_vertex_weight = true;
        command->vw = w;
      }
    } else if (extra_components == 3) {
      // v x y z r g b  (6 total) — color, weight = r (legacy compat)
      real_t cr = r, cg = g, cb = b;
      const char *color_cursor = extra_token;
      if (opt_tryParseFloatToken(&cr, &color_cursor) &&
          opt_tryParseFloatToken(&cg, &color_cursor) &&
          opt_tryParseFloatToken(&cb, &color_cursor)) {
        command->has_vertex_weight = true;
        command->vw = cr;
        command->has_vertex_color = true;
        command->vc_r = cr;
        command->vc_g = cg;
        command->vc_b = cb;
      }
    } else if (extra_components >= 4) {
      // v x y z w r g b  (7+ total) — weight + color.
      // NOTE: 7-component vertices are NOT part of the OBJ spec nor the common
      // vertex-color extension, which define only xyz / xyzw / xyzrgb and treat
      // weight and color as mutually exclusive.  tinyobjloader accepts this as
      // an extension: the 4th value is the weight, the next three are RGB.
      // (The classic LoadObj / LoadObjWithCallback path instead caps at 6 and
      // ignores any extra tokens, so it differs from LoadObjOpt here.)
      real_t vw = real_t(1.0), cr = r, cg = g, cb = b;
      const char *wc_cursor = extra_token;
      if (opt_tryParseFloatToken(&vw, &wc_cursor) &&
          opt_tryParseFloatToken(&cr, &wc_cursor) &&
          opt_tryParseFloatToken(&cg, &wc_cursor) &&
          opt_tryParseFloatToken(&cb, &wc_cursor)) {
        command->has_vertex_weight = true;
        command->vw = vw;
        command->has_vertex_color = true;
        command->vc_r = cr;
        command->vc_g = cg;
        command->vc_b = cb;
      }
    }
    command->vx = x;
    command->vy = y;
    command->vz = z;
    command->type = OPT_CMD_V;
    return true;
  }

  // normal
  if (token[0] == 'v' && token[1] == 'n' && TINYOBJ_OPT_IS_SPACE(token[2])) {
    token += 3;
    real_t x = 0, y = 0, z = 0;
    if (!opt_tryParseFloatToken(&x, &token) ||
        !opt_tryParseFloatToken(&y, &token) ||
        !opt_tryParseFloatToken(&z, &token)) {
      if (parse_error) *parse_error = true;
      if (parse_error_message) *parse_error_message = "failed to parse `vn' line";
      return false;
    }
    command->nx = x;
    command->ny = y;
    command->nz = z;
    command->type = OPT_CMD_VN;
    return true;
  }

  // texcoord
  if (token[0] == 'v' && token[1] == 't' && TINYOBJ_OPT_IS_SPACE(token[2])) {
    token += 3;
    real_t x = 0, y = 0, w = 0;
    if (!opt_tryParseFloatToken(&x, &token)) {
      if (parse_error) *parse_error = true;
      if (parse_error_message) *parse_error_message = "failed to parse `vt' line";
      return false;
    }
    const char *y_token = token;
    if (opt_tryParseFloatToken(&y, &y_token)) {
      token = y_token;
    }
    const char *extra_token = token;
    opt_skip_space(&extra_token);
    if (!TINYOBJ_OPT_IS_NEW_LINE(extra_token[0]) && extra_token[0] != '#' &&
        opt_tryParseFloatToken(&w, &extra_token)) {
      command->has_texcoord_w = true;
      command->tw = w;
    }
    command->tx = x;
    command->ty = y;
    command->type = OPT_CMD_VT;
    return true;
  }

  // face
  if (token[0] == 'f' && TINYOBJ_OPT_IS_SPACE(token[1])) {
    token += 2;
    opt_skip_space(&token);

    // Collect face vertices (use small stack buffer for typical case)
    opt_index_t face_buf[8];
    std::vector<opt_index_t> face_overflow;
    int face_count = 0;

    while (!TINYOBJ_OPT_IS_NEW_LINE(token[0]) && !opt_is_comment_start(token)) {
      opt_index_t vi;
      if (!opt_parseRawTriple(&token, &vi)) {
        if (parse_error) *parse_error = true;
        if (parse_error_message) {
          *parse_error_message = "failed to parse `f' line (invalid vertex index)";
        }
        return false;
      }
      opt_skip_space(&token);
      if (face_count < 8) {
        face_buf[face_count++] = vi;
      } else {
        if (face_count == 8) {
          face_overflow.reserve(16);
          for (int k = 0; k < 8; k++) face_overflow.push_back(face_buf[k]);
        }
        face_overflow.push_back(vi);
        face_count++;
      }
    }

      command->type = OPT_CMD_F;
      command->face_vertex_count = static_cast<unsigned int>(face_count);

    // Preserve degenerate faces so warnings and EOF shape behavior can match
    // the legacy loader, but do not reserve output slots for them.
    if (face_count < 3) {
      command->degenerate_face = true;
      command->emitted_face_count = 0;
      command->emitted_face_verts = 0;
      command->f_count = 0;
      return true;
    }

      if (triangulate) {
        command->emitted_face_count = static_cast<unsigned int>(face_count - 2);
        command->emitted_face_verts = 3;
        command->f_count = command->emitted_face_count * 3;

      opt_index_t *dst = command->f_inline;
      if (command->face_vertex_count > OptCommand::kInlineIndexCapacity) {
        command->f_heap.resize(command->face_vertex_count);
        dst = command->f_heap.data();
      }
      if (face_count <= 8) {
        for (int k = 0; k < face_count; k++) dst[k] = face_buf[k];
      } else {
        for (size_t k = 0; k < face_overflow.size(); k++) dst[k] = face_overflow[k];
      }
    } else {
      command->emitted_face_count = 1;
      command->emitted_face_verts = static_cast<unsigned int>(face_count);
      command->f_count = static_cast<unsigned int>(face_count);
      command->face_vertex_count = static_cast<unsigned int>(face_count);

      opt_index_t *dst = command->f_inline;
      if (command->f_count > OptCommand::kInlineIndexCapacity) {
        command->f_heap.resize(command->f_count);
        dst = command->f_heap.data();
      }

      if (face_count <= 8) {
        for (int k = 0; k < face_count; k++) dst[k] = face_buf[k];
      } else {
        for (size_t k = 0; k < face_overflow.size(); k++) {
          dst[k] = face_overflow[k];
        }
      }
    }
    return true;
  }

  // usemtl
  if (std::strncmp(token, "usemtl", 6) == 0 &&
      TINYOBJ_OPT_IS_SPACE(token[6])) {
    token += 7;
    opt_skip_space(&token);
    command->material_name = token;
    command->material_name_len = static_cast<unsigned int>(
        opt_length_until_token_or_comment(token,
                                p_len - static_cast<size_t>(token - p)));
    command->type = OPT_CMD_USEMTL;
    return true;
  }

  // mtllib
  if (std::strncmp(token, "mtllib", 6) == 0 &&
      TINYOBJ_OPT_IS_SPACE(token[6])) {
    token += 7;
    opt_skip_space(&token);
    command->mtllib_name = token;
    command->mtllib_name_len = static_cast<unsigned int>(
        opt_length_until_newline(token,
                                p_len - static_cast<size_t>(token - p)));
    command->type = OPT_CMD_MTLLIB;
    return true;
  }

  // group
  if (token[0] == 'g' &&
      (TINYOBJ_OPT_IS_SPACE(token[1]) || TINYOBJ_OPT_IS_NEW_LINE(token[1]) ||
       token[1] == '\0')) {
    if (TINYOBJ_OPT_IS_SPACE(token[1])) {
      token += 2;
    } else {
      token += 1;
    }
    command->group_name_storage = opt_parseGroupName(
        token, p_len - static_cast<size_t>(token - p));
    command->group_name = nullptr;
    command->group_name_len =
        static_cast<unsigned int>(command->group_name_storage.size());
    command->group_name_empty = command->group_name_storage.empty();
    command->type = OPT_CMD_G;
    return true;
  }

  // object
  if (token[0] == 'o' && TINYOBJ_OPT_IS_SPACE(token[1])) {
    token += 2;
    command->object_name = token;
    command->object_name_len = static_cast<unsigned int>(
        p_len - static_cast<size_t>(token - p));
    command->type = OPT_CMD_O;
    return true;
  }

  // smoothing group
  if (token[0] == 's' && TINYOBJ_OPT_IS_SPACE(token[1])) {
    token += 2;
    opt_skip_space(&token);
    if (TINYOBJ_OPT_IS_NEW_LINE(token[0])) {
      command->smoothing_group_id = 0;
    } else if (token[0] == 'o' && token[1] == 'f' && token[2] == 'f' &&
               (TINYOBJ_OPT_IS_NEW_LINE(token[3]) || token[3] == '\0' ||
                token[3] == ' ' || token[3] == '\t')) {
      command->smoothing_group_id = 0;
    } else {
      const int sm = opt_my_atoi(token);
      command->smoothing_group_id = (sm > 0) ? static_cast<unsigned int>(sm) : 0;
    }
    command->type = OPT_CMD_S;
    return true;
  }

  return false;
}

// ---- SIMD newline scanning ----

#ifdef TINYOBJLOADER_USE_SIMD

// Include <intrin.h> for _BitScanForward on MSVC. Placed here (not at the top)
// because it's only needed when TINYOBJLOADER_USE_SIMD is enabled.
#if defined(_MSC_VER)
#include <intrin.h>
#endif

// Portable count-trailing-zeros for SIMD bitmask extraction
static inline unsigned int tinyobj_ctz(unsigned int x) {
#if defined(_MSC_VER)
  unsigned long idx;
  _BitScanForward(&idx, x);
  return static_cast<unsigned int>(idx);
#else
  return static_cast<unsigned int>(__builtin_ctz(x));
#endif
}

#if defined(TINYOBJLOADER_SIMD_AVX2)

/// AVX2-accelerated line-ending scanning — finds '\n' and '\r' positions.
static void simd_find_newlines(const char *buf, size_t len,
                               std::vector<size_t> &positions) {
  const __m256i nl = _mm256_set1_epi8('\n');
  const __m256i cr = _mm256_set1_epi8('\r');
  size_t i = 0;
  for (; i + 32 <= len; i += 32) {
    __m256i chunk =
        _mm256_loadu_si256(reinterpret_cast<const __m256i *>(buf + i));
    __m256i cmp_nl = _mm256_cmpeq_epi8(chunk, nl);
    __m256i cmp_cr = _mm256_cmpeq_epi8(chunk, cr);
    __m256i combined = _mm256_or_si256(cmp_nl, cmp_cr);
    unsigned int mask = static_cast<unsigned int>(_mm256_movemask_epi8(combined));
    while (mask) {
      unsigned int bit = tinyobj_ctz(mask);
      positions.push_back(i + bit);
      mask &= mask - 1;
    }
  }
  // Scalar tail
  for (; i < len; i++) {
    if (buf[i] == '\n' || buf[i] == '\r') positions.push_back(i);
  }
}

#elif defined(TINYOBJLOADER_SIMD_SSE2)

/// SSE2-accelerated line-ending scanning — finds '\n' and '\r' positions.
static void simd_find_newlines(const char *buf, size_t len,
                               std::vector<size_t> &positions) {
  const __m128i nl = _mm_set1_epi8('\n');
  const __m128i cr = _mm_set1_epi8('\r');
  size_t i = 0;
  for (; i + 16 <= len; i += 16) {
    __m128i chunk =
        _mm_loadu_si128(reinterpret_cast<const __m128i *>(buf + i));
    __m128i cmp_nl = _mm_cmpeq_epi8(chunk, nl);
    __m128i cmp_cr = _mm_cmpeq_epi8(chunk, cr);
    __m128i combined = _mm_or_si128(cmp_nl, cmp_cr);
    int mask = _mm_movemask_epi8(combined);
    while (mask) {
      int bit = static_cast<int>(tinyobj_ctz(static_cast<unsigned int>(mask)));
      positions.push_back(i + static_cast<size_t>(bit));
      mask &= mask - 1;
    }
  }
  // Scalar tail
  for (; i < len; i++) {
    if (buf[i] == '\n' || buf[i] == '\r') positions.push_back(i);
  }
}

#elif defined(TINYOBJLOADER_SIMD_NEON)

/// NEON-accelerated line-ending scanning — finds '\n' and '\r' positions.
static void simd_find_newlines(const char *buf, size_t len,
                               std::vector<size_t> &positions) {
  const uint8x16_t nl = vdupq_n_u8('\n');
  const uint8x16_t cr = vdupq_n_u8('\r');
  size_t i = 0;
  for (; i + 16 <= len; i += 16) {
    uint8x16_t chunk = vld1q_u8(reinterpret_cast<const uint8_t *>(buf + i));
    uint8x16_t cmp_nl = vceqq_u8(chunk, nl);
    uint8x16_t cmp_cr = vceqq_u8(chunk, cr);
    uint8x16_t combined = vorrq_u8(cmp_nl, cmp_cr);
    for (int j = 0; j < 16; j++) {
      if (vgetq_lane_u8(combined, 0) != 0) {
        positions.push_back(i + static_cast<size_t>(j));
      }
      combined = vextq_u8(combined, combined, 1);
    }
  }
  for (; i < len; i++) {
    if (buf[i] == '\n' || buf[i] == '\r') positions.push_back(i);
  }
}

#endif  // SIMD variant

/// Build LineInfo array from SIMD-detected line-ending positions.
/// The positions array may contain both '\r' and '\n' hits;
/// consecutive \r\n pairs are collapsed into a single line boundary.
static void simd_build_line_infos(const char *buf, size_t len,
                                  const std::vector<size_t> &nl_positions,
                                  std::vector<LineInfo> &out) {
  out.reserve(nl_positions.size() + 1);
  size_t prev = 0;
  for (size_t k = 0; k < nl_positions.size(); k++) {
    size_t pos = nl_positions[k];
    // Skip the '\n' in a '\r\n' pair (the '\r' already created a boundary)
    if (buf[pos] == '\n' && pos > 0 && buf[pos - 1] == '\r') {
      prev = pos + 1;
      continue;
    }
    size_t line_len = pos - prev;
    if (line_len > 0) {
      LineInfo info;
      info.pos = prev;
      info.len = line_len;
      out.push_back(info);
    }
    prev = pos + 1;
  }
  // Handle last line without trailing newline
  if (prev < len) {
    size_t line_len = len - prev;
    if (line_len > 0) {
      LineInfo info;
      info.pos = prev;
      info.len = line_len;
      out.push_back(info);
    }
  }
}

#endif  // TINYOBJLOADER_USE_SIMD

/// Scalar fallback newline scanning — handles \n, \r\n, and bare \r
static void scalar_find_line_infos(const char *buf, size_t start, size_t end,
                                   std::vector<LineInfo> &out) {
  size_t prev = start;
  for (size_t i = start; i < end; i++) {
    if (buf[i] == '\n' || buf[i] == '\r') {
      size_t line_len = i - prev;
      // Skip \r in \r\n pair
      if (buf[i] == '\r' && (i + 1 < end) && buf[i + 1] == '\n') {
        if (line_len > 0) {
          LineInfo info;
          info.pos = prev;
          info.len = line_len;
          out.push_back(info);
        }
        i++;  // skip the \n in \r\n
      } else {
        // bare \n or bare \r
        if (line_len > 0) {
          LineInfo info;
          info.pos = prev;
          info.len = line_len;
          out.push_back(info);
        }
      }
      prev = i + 1;
    }
  }
  if (prev < end) {
    size_t line_len = end - prev;
    if (line_len > 0) {
      LineInfo info;
      info.pos = prev;
      info.len = line_len;
      out.push_back(info);
    }
  }
}

template <typename DstVec, typename SrcVec>
static void opt_assign_vector(DstVec *dst, const SrcVec &src) {
  if (!dst) return;
  dst->assign(src.begin(), src.end());
}

static void ConvertLegacyShapeToBasic(const shape_t &src, basic_shape_t<> *dst) {
  if (!dst) return;
  dst->name = src.name;
  opt_assign_vector(&dst->mesh.indices, src.mesh.indices);
  opt_assign_vector(&dst->mesh.num_face_vertices, src.mesh.num_face_vertices);
  opt_assign_vector(&dst->mesh.material_ids, src.mesh.material_ids);
  opt_assign_vector(&dst->mesh.smoothing_group_ids,
                    src.mesh.smoothing_group_ids);
  dst->mesh.tags = src.mesh.tags;
  opt_assign_vector(&dst->lines.indices, src.lines.indices);
  opt_assign_vector(&dst->lines.num_line_vertices, src.lines.num_line_vertices);
  opt_assign_vector(&dst->points.indices, src.points.indices);
}

static void ConvertLegacyResultToBasic(
    const attrib_t &src_attrib, const std::vector<shape_t> &src_shapes,
    const std::vector<material_t> &src_materials, basic_attrib_t<> *dst_attrib,
    std::vector<basic_shape_t<> > *dst_shapes,
    std::vector<material_t> *dst_materials) {
  if (dst_attrib) {
    opt_assign_vector(&dst_attrib->vertices, src_attrib.vertices);
    opt_assign_vector(&dst_attrib->vertex_weights, src_attrib.vertex_weights);
    opt_assign_vector(&dst_attrib->normals, src_attrib.normals);
    opt_assign_vector(&dst_attrib->texcoords, src_attrib.texcoords);
    opt_assign_vector(&dst_attrib->texcoord_ws, src_attrib.texcoord_ws);
    opt_assign_vector(&dst_attrib->colors, src_attrib.colors);
    dst_attrib->skin_weights = src_attrib.skin_weights;
    dst_attrib->indices.clear();
    dst_attrib->face_num_verts.clear();
    dst_attrib->material_ids.clear();
    for (size_t si = 0; si < src_shapes.size(); si++) {
      const mesh_t &mesh = src_shapes[si].mesh;
      dst_attrib->indices.insert(dst_attrib->indices.end(), mesh.indices.begin(),
                                 mesh.indices.end());
      dst_attrib->face_num_verts.insert(dst_attrib->face_num_verts.end(),
                                        mesh.num_face_vertices.begin(),
                                        mesh.num_face_vertices.end());
      dst_attrib->material_ids.insert(dst_attrib->material_ids.end(),
                                      mesh.material_ids.begin(),
                                      mesh.material_ids.end());
    }
  }

  if (dst_shapes) {
    dst_shapes->clear();
    dst_shapes->resize(src_shapes.size());
    for (size_t i = 0; i < src_shapes.size(); i++) {
      ConvertLegacyShapeToBasic(src_shapes[i], &(*dst_shapes)[i]);
    }
  }

  if (dst_materials) {
    (*dst_materials) = src_materials;
  }
}

/// Check if the first non-whitespace token on a line (starting at p, length
/// rem) requires the legacy parser.
static inline bool opt_line_start_is_legacy(const char *p, size_t rem) {
  while (rem > 0 && (*p == ' ' || *p == '\t')) { p++; rem--; }
  if (rem >= 3 && p[0] == 'v' && p[1] == 'w' &&
      (p[2] == ' ' || p[2] == '\t'))
    return true;
  if (rem >= 2 &&
      ((p[0] == 'l' || p[0] == 'p' || p[0] == 't') &&
       (p[1] == ' ' || p[1] == '\t')))
    return true;
  return false;
}

/// Fast whole-buffer scan for legacy-only tokens (vw, l, p, t at line start).
/// Uses memchr for SIMD-accelerated newline finding.  Handles both \n and
/// bare \r (old Mac) line endings.
static bool opt_buffer_requires_legacy_fallback(const char *buf, size_t len) {
  if (!buf || len == 0) return false;
  // Check at start of buffer
  if (opt_line_start_is_legacy(buf, len)) return true;
  // Scan for line breaks (\n first, then bare \r)
  for (int pass = 0; pass < 2; pass++) {
    const char needle = (pass == 0) ? '\n' : '\r';
    const char *p = buf;
    size_t remaining = len;
    for (;;) {
      const char *nl = static_cast<const char *>(
          std::memchr(p, needle, remaining));
      if (!nl) break;
      size_t offset = static_cast<size_t>(nl - buf) + 1;
      if (offset >= len) break;
      // Skip \n in \r\n pair on the \r pass to avoid double-checking
      if (needle == '\r' && offset < len && buf[offset] == '\n') {
        p = buf + offset + 1;
        remaining = len - offset - 1;
        continue;
      }
      if (opt_line_start_is_legacy(buf + offset, len - offset)) return true;
      p = buf + offset;
      remaining = len - offset;
    }
  }
  return false;
}

///
/// Trie-based string → float cache for repeated OBJ coordinate values.
/// Caches short float tokens (up to kMaxKeyLen chars) in a compact trie.
/// Alphabet: '0'-'9', '+', '-', '.', 'e', 'E' (15 chars).
/// Thread-local — each thread gets its own cache instance.
///
class OptFloatCache {
 public:
  static const int kFp32MaxKeyLen =
      std::numeric_limits<float>::digits10 + 2;  // 8
  static const int kRealMaxKeyLen =
      std::numeric_limits<real_t>::digits10 + 2;
  static const int kAlphabetSize = 15;

  explicit OptFloatCache(int max_nodes = 1024, bool fp32_keys = true)
      : max_nodes_(max_nodes < 2 ? 2 : (max_nodes > 65535 ? 65535 : max_nodes)),
        max_key_len_(fp32_keys ? kFp32MaxKeyLen : kRealMaxKeyLen) {
    nodes_.reserve(static_cast<size_t>(max_nodes_));
    nodes_.resize(1);
    std::memset(&nodes_[0], 0, sizeof(Node));
  }

  static inline int char_index(char c) {
    if (c >= '0' && c <= '9') return c - '0';
    switch (c) {
      case '+': return 10;
      case '-': return 11;
      case '.': return 12;
      case 'e': return 13;
      case 'E': return 14;
    }
    return -1;
  }

  /// Try cache lookup.  Returns true + sets *out on hit.
  inline bool lookup(const char *key, int key_len, real_t *out) const {
    if (key_len <= 0 || key_len > max_key_len_) return false;
    int node = 0;
    for (int i = 0; i < key_len; i++) {
      int ci = char_index(key[i]);
      if (ci < 0) return false;
      int child = nodes_[static_cast<size_t>(node)]
                      .children[static_cast<size_t>(ci)];
      if (child == 0) return false;
      node = child;
    }
    const Node &n = nodes_[static_cast<size_t>(node)];
    if (!n.has_value) return false;
    *out = n.value;
    return true;
  }

  /// Insert parsed value.  Returns false if full or key too long.
  inline bool insert(const char *key, int key_len, real_t value) {
    if (key_len <= 0 || key_len > max_key_len_) return false;
    int node = 0;
    for (int i = 0; i < key_len; i++) {
      int ci = char_index(key[i]);
      if (ci < 0) return false;
      uint16_t child = nodes_[static_cast<size_t>(node)]
                           .children[static_cast<size_t>(ci)];
      if (child == 0) {
        if (static_cast<int>(nodes_.size()) >= max_nodes_) return false;
        child = static_cast<uint16_t>(nodes_.size());
        // Write back BEFORE push_back (push_back may realloc)
        nodes_[static_cast<size_t>(node)]
            .children[static_cast<size_t>(ci)] = child;
        Node new_node;
        std::memset(&new_node, 0, sizeof(Node));
        nodes_.push_back(new_node);
      }
      node = static_cast<int>(child);
    }
    Node &n = nodes_[static_cast<size_t>(node)];
    n.value = value;
    n.has_value = true;
    return true;
  }

  int max_key_len() const { return max_key_len_; }

 private:
  struct Node {
    uint16_t children[kAlphabetSize];  // 30 bytes
    real_t value;                      // 4 or 8 bytes
    bool has_value;                    // 1 byte
  };
  std::vector<Node> nodes_;
  int max_nodes_;
  int max_key_len_;
};

// Fast inline float parse: skip whitespace, call fast_float with line_end,
// advance cursor.  No token-end pre-scan.  Returns false if no float found.
// When cache is non-null, checks/populates the trie cache for short tokens.
// Handles nan/inf keywords with OBJ-compatible replacement values (matching
// opt_tryParseDouble behavior).
#ifndef TINYOBJLOADER_DISABLE_FAST_FLOAT
static inline bool opt_fast_parse_float(real_t *out, const char **cursor,
                                        const char *line_end,
                                        OptFloatCache *cache = nullptr) {
  const char *p = *cursor;
  while (p < line_end && (*p == ' ' || *p == '\t')) p++;
  if (p >= line_end || *p == '\n' || *p == '\r' || *p == '#') return false;

  // Check for nan/inf keywords before calling fast_float (which doesn't
  // handle them).  Map to OBJ-compatible values: nan→0, +inf→max, -inf→lowest.
  {
    const char *q = p;
    if (q < line_end && (*q == '+' || *q == '-')) ++q;
    if (q < line_end) {
      char fc = *q;
      if (fc >= 'A' && fc <= 'Z') fc += 32;
      if (fc == 'n' || fc == 'i') {
        double special_val;
        const char *end_ptr;
        if (detail_fp::tryParseNanInf(p, line_end, &special_val, &end_ptr)) {
          *out = static_cast<real_t>(special_val);
          *cursor = end_ptr;
          return true;
        }
      }
    }
  }

  // --- Cache fast path: scan up to max_key_len chars for trie lookup ---
  if (cache) {
    const char *tok_start = p;
    const char *kp = p;
    int klen = 0;
    int mkl = cache->max_key_len();
    while (klen < mkl && kp < line_end) {
      char c = *kp;
      if (c == ' ' || c == '\t' || c == '\n' || c == '\r' || c == '#' ||
          c == '\0')
        break;
      if (OptFloatCache::char_index(c) < 0) break;
      klen++;
      kp++;
    }
    // Token is cacheable if it ended at a delimiter (not mid-token)
    bool at_delim = (kp >= line_end || *kp == ' ' || *kp == '\t' ||
                     *kp == '\n' || *kp == '\r' || *kp == '#' || *kp == '\0');
    if (klen > 0 && at_delim) {
      real_t cached;
      if (cache->lookup(tok_start, klen, &cached)) {
        *out = cached;
        *cursor = kp;
        return true;
      }
      // Cache miss — parse normally, then insert
      real_t tmp;
      auto r = fast_float::from_chars(tok_start, line_end, tmp,
          fast_float::chars_format::general |
          fast_float::chars_format::allow_leading_plus);
      if (r.ec != tinyobj_ff::ff_errc::ok) return false;
      *out = tmp;
      *cursor = r.ptr;
      cache->insert(tok_start, klen, tmp);
      return true;
    }
    // Token too long or has non-float chars — fall through to uncached parse
  }

  // --- Uncached parse ---
  real_t tmp;
  auto r = fast_float::from_chars(p, line_end, tmp,
      fast_float::chars_format::general |
      fast_float::chars_format::allow_leading_plus);
  if (r.ec != tinyobj_ff::ff_errc::ok) return false;
  *out = tmp;
  *cursor = r.ptr;
  return true;
}
#endif

// Parse a single line and store into compact per-type thread data.
// Fast path: v/vn/vt lines are parsed directly using fast_float with line_end,
// bypassing OptCommand entirely.  Face/meta lines fall back to opt_parseLine.
static void opt_parseLineToThreadData(
    OptThreadData &td, const char *line_ptr, size_t line_len,
    bool triangulate, size_t line_number,
    OptFloatCache *cache = nullptr) {
  const char *token = line_ptr;
  while (*token == ' ' || *token == '\t') token++;
  if (*token == '\n' || *token == '\r' || *token == '#' || *token == '\0')
    return;

#ifdef TINYOBJLOADER_DISABLE_FAST_FLOAT
  (void)cache;  // only consumed by the fast_float fast path below
#else
  const char *line_end = line_ptr + line_len;

  // ---- Fast path: vertex position (v x y z [w] [r g b]) ----
  if (token[0] == 'v' && (token[1] == ' ' || token[1] == '\t')) {
    const char *cur = token + 2;
    real_t x, y, z;
    if (!opt_fast_parse_float(&x, &cur, line_end, cache) ||
        !opt_fast_parse_float(&y, &cur, line_end, cache) ||
        !opt_fast_parse_float(&z, &cur, line_end, cache)) {
      td.error_line = line_number;
      td.error_message = "failed to parse `v' line";
      return;
    }
    // Write xyz (batch: resize + direct write)
    size_t off = td.v_pos.size();
    td.v_pos.resize(off + 3);
    td.v_pos[off] = x; td.v_pos[off+1] = y; td.v_pos[off+2] = z;

    // Check for extra components (weight / color)
    real_t extra[4];
    int nextra = 0;
    while (nextra < 4 && opt_fast_parse_float(&extra[nextra], &cur, line_end, cache))
      nextra++;

    if (nextra == 1) {
      // v x y z w  (4 total) — weight only, no color
      if (!td.saw_any_weight) {
        td.saw_any_weight = true;
        td.v_weight.resize(td.num_v, real_t(1.0));
      }
      td.v_weight.push_back(extra[0]);
      td.saw_missing_color = true;
      if (td.saw_any_color) {
        size_t coff = td.v_color.size();
        td.v_color.resize(coff + 3);
        td.v_color[coff] = real_t(1.0);
        td.v_color[coff+1] = real_t(1.0);
        td.v_color[coff+2] = real_t(1.0);
      }
    } else if (nextra == 3) {
      // v x y z r g b  (6 total) — color, weight = r (legacy compat)
      if (!td.saw_any_weight) {
        td.saw_any_weight = true;
        td.v_weight.resize(td.num_v, real_t(1.0));
      }
      td.v_weight.push_back(extra[0]);
      if (!td.saw_any_color) {
        td.saw_any_color = true;
        td.v_color.resize(td.num_v * 3, real_t(1.0));
      }
      size_t coff = td.v_color.size();
      td.v_color.resize(coff + 3);
      td.v_color[coff] = extra[0];
      td.v_color[coff+1] = extra[1];
      td.v_color[coff+2] = extra[2];
    } else if (nextra >= 4) {
      // v x y z w r g b  (7+ total) — weight + color.
      // NOTE: 7-component vertices are NOT part of the OBJ spec nor the common
      // vertex-color extension, which define only xyz / xyzw / xyzrgb and treat
      // weight and color as mutually exclusive.  tinyobjloader accepts this as
      // an extension: the 4th value is the weight, the next three are RGB.
      // (The classic LoadObj / LoadObjWithCallback path instead caps at 6 and
      // ignores any extra tokens, so it differs from LoadObjOpt here.)
      if (!td.saw_any_weight) {
        td.saw_any_weight = true;
        td.v_weight.resize(td.num_v, real_t(1.0));
      }
      td.v_weight.push_back(extra[0]);
      if (!td.saw_any_color) {
        td.saw_any_color = true;
        td.v_color.resize(td.num_v * 3, real_t(1.0));
      }
      size_t coff = td.v_color.size();
      td.v_color.resize(coff + 3);
      td.v_color[coff] = extra[1];
      td.v_color[coff+1] = extra[2];
      td.v_color[coff+2] = extra[3];
    } else {
      // nextra == 0 or 2 — no color
      if (td.saw_any_weight) td.v_weight.push_back(real_t(1.0));
      td.saw_missing_color = true;
      if (td.saw_any_color) {
        size_t coff = td.v_color.size();
        td.v_color.resize(coff + 3);
        td.v_color[coff] = real_t(1.0);
        td.v_color[coff+1] = real_t(1.0);
        td.v_color[coff+2] = real_t(1.0);
      }
    }
    td.num_v++;
    return;
  }

  // ---- Fast path: normal (vn x y z) ----
  if (token[0] == 'v' && token[1] == 'n' &&
      (token[2] == ' ' || token[2] == '\t')) {
    const char *cur = token + 3;
    real_t x, y, z;
    if (!opt_fast_parse_float(&x, &cur, line_end, cache) ||
        !opt_fast_parse_float(&y, &cur, line_end, cache) ||
        !opt_fast_parse_float(&z, &cur, line_end, cache)) {
      td.error_line = line_number;
      td.error_message = "failed to parse `vn' line";
      return;
    }
    size_t off = td.vn_data.size();
    td.vn_data.resize(off + 3);
    td.vn_data[off] = x; td.vn_data[off+1] = y; td.vn_data[off+2] = z;
    td.num_vn++;
    return;
  }

  // ---- Fast path: texcoord (vt u [v [w]]) ----
  if (token[0] == 'v' && token[1] == 't' &&
      (token[2] == ' ' || token[2] == '\t')) {
    const char *cur = token + 3;
    real_t u = 0, v = 0;
    if (!opt_fast_parse_float(&u, &cur, line_end, cache)) {
      td.error_line = line_number;
      td.error_message = "failed to parse `vt' line";
      return;
    }
    opt_fast_parse_float(&v, &cur, line_end, cache);  // v is optional
    real_t w = 0;
    bool has_w = opt_fast_parse_float(&w, &cur, line_end, cache);

    size_t off = td.vt_data.size();
    td.vt_data.resize(off + 2);
    td.vt_data[off] = u; td.vt_data[off+1] = v;
    if (has_w) {
      if (!td.saw_any_texcoord_w) {
        td.saw_any_texcoord_w = true;
        td.vt_w.resize(td.num_vt, real_t(0.0));
      }
      td.vt_w.push_back(w);
    } else if (td.saw_any_texcoord_w) {
      td.vt_w.push_back(real_t(0.0));
    }
    td.num_vt++;
    return;
  }
#endif  // TINYOBJLOADER_DISABLE_FAST_FLOAT

  // ---- Slow path: face / meta commands via opt_parseLine ----
  OptCommand cmd;
  bool parse_error = false;
  bool ok = opt_parseLine(&cmd, line_ptr, line_len, triangulate,
                           &parse_error, nullptr);
  if (parse_error) {
    td.error_line = line_number;
    std::string msg;
    opt_parseLine(&cmd, line_ptr, line_len, triangulate, nullptr, &msg);
    td.error_message = msg;
    return;
  }
  if (!ok) return;

  switch (cmd.type) {
#ifdef TINYOBJLOADER_DISABLE_FAST_FLOAT
    // When fast_float is disabled, v/vn/vt fall through to here
    case OPT_CMD_V: {
      size_t off = td.v_pos.size();
      td.v_pos.resize(off + 3);
      td.v_pos[off] = cmd.vx; td.v_pos[off+1] = cmd.vy; td.v_pos[off+2] = cmd.vz;
      if (cmd.has_vertex_weight) {
        if (!td.saw_any_weight) {
          td.saw_any_weight = true;
          td.v_weight.resize(td.num_v, real_t(1.0));
        }
        td.v_weight.push_back(cmd.vw);
      } else if (td.saw_any_weight) {
        td.v_weight.push_back(real_t(1.0));
      }
      if (cmd.has_vertex_color) {
        if (!td.saw_any_color) {
          td.saw_any_color = true;
          td.v_color.resize(td.num_v * 3, real_t(1.0));
        }
        td.v_color.push_back(cmd.vc_r);
        td.v_color.push_back(cmd.vc_g);
        td.v_color.push_back(cmd.vc_b);
      } else {
        td.saw_missing_color = true;
        if (td.saw_any_color) {
          td.v_color.push_back(real_t(1.0));
          td.v_color.push_back(real_t(1.0));
          td.v_color.push_back(real_t(1.0));
        }
      }
      td.num_v++;
      break;
    }
    case OPT_CMD_VN: {
      size_t off = td.vn_data.size();
      td.vn_data.resize(off + 3);
      td.vn_data[off] = cmd.nx; td.vn_data[off+1] = cmd.ny; td.vn_data[off+2] = cmd.nz;
      td.num_vn++;
      break;
    }
    case OPT_CMD_VT: {
      size_t off = td.vt_data.size();
      td.vt_data.resize(off + 2);
      td.vt_data[off] = cmd.tx; td.vt_data[off+1] = cmd.ty;
      if (cmd.has_texcoord_w) {
        if (!td.saw_any_texcoord_w) {
          td.saw_any_texcoord_w = true;
          td.vt_w.resize(td.num_vt, real_t(0.0));
        }
        td.vt_w.push_back(cmd.tw);
      } else if (td.saw_any_texcoord_w) {
        td.vt_w.push_back(real_t(0.0));
      }
      td.num_vt++;
      break;
    }
#endif  // TINYOBJLOADER_DISABLE_FAST_FLOAT
    case OPT_CMD_F: {
      OptFaceCmd fc;
      fc.face_vertex_count = cmd.face_vertex_count;
      fc.emitted_face_count = cmd.emitted_face_count;
      fc.emitted_face_verts = cmd.emitted_face_verts;
      fc.f_count = cmd.f_count;
      fc.degenerate = cmd.degenerate_face;
      fc.source_line = static_cast<uint32_t>(line_number);
      fc.v_count_before = static_cast<uint32_t>(td.num_v);
      fc.vn_count_before = static_cast<uint32_t>(td.num_vn);
      fc.vt_count_before = static_cast<uint32_t>(td.num_vt);
      if (!cmd.f_heap.empty()) {
        fc.f_heap = std::move(cmd.f_heap);
      } else if (cmd.face_vertex_count <= OptFaceCmd::kInlineCap) {
        for (uint32_t k = 0; k < cmd.face_vertex_count; k++)
          fc.f_inline[k] = cmd.f_inline[k];
      } else {
        fc.f_heap.assign(cmd.f_inline, cmd.f_inline + cmd.face_vertex_count);
      }
      td.num_f_indices += fc.f_count;
      td.num_f_faces += fc.emitted_face_count;
      OptSeqEntry se;
      se.kind = OptSeqEntry::SEQ_FACE;
      se.index = static_cast<uint32_t>(td.faces.size());
      td.seq.push_back(se);
      td.faces.push_back(std::move(fc));
      break;
    }
    case OPT_CMD_G:
    case OPT_CMD_O:
    case OPT_CMD_USEMTL:
    case OPT_CMD_MTLLIB:
    case OPT_CMD_S: {
      OptMetaCmd mc;
      mc.type = cmd.type;
      mc.source_line = static_cast<uint32_t>(line_number);
      if (cmd.type == OPT_CMD_G) {
        mc.str_storage = cmd.group_name_storage;
        mc.str_ptr = cmd.group_name;
        mc.str_len = cmd.group_name_len;
        mc.group_name_empty = cmd.group_name_empty;
      } else if (cmd.type == OPT_CMD_O) {
        mc.str_ptr = cmd.object_name;
        mc.str_len = cmd.object_name_len;
      } else if (cmd.type == OPT_CMD_USEMTL) {
        mc.str_ptr = cmd.material_name;
        mc.str_len = cmd.material_name_len;
      } else if (cmd.type == OPT_CMD_MTLLIB) {
        mc.str_ptr = cmd.mtllib_name;
        mc.str_len = cmd.mtllib_name_len;
      } else if (cmd.type == OPT_CMD_S) {
        mc.smoothing_group_id = cmd.smoothing_group_id;
        td.saw_any_smoothing = true;
      }
      OptSeqEntry se;
      se.kind = OptSeqEntry::SEQ_META;
      se.index = static_cast<uint32_t>(td.metas.size());
      td.seq.push_back(se);
      td.metas.push_back(std::move(mc));
      break;
    }
    default:
      break;
  }
}

}  // namespace opt_internal

// Internal implementation with optional basedir for material path resolution
static bool LoadObjOpt_internal(basic_attrib_t<> *attrib,
                                std::vector<basic_shape_t<>> *shapes,
                                std::vector<material_t> *materials,
                                std::string *warn, std::string *err,
                                const char *buf, size_t buf_len,
                                const std::string &mtl_basedir,
                                const std::string &source_name,
                                bool enable_mtllib_loading,
                                const OptLoadConfig &config) {
  using namespace opt_internal;

  if (!attrib || !shapes) {
    if (err) *err = "attrib and shapes must not be null.";
    return false;
  }

  attrib->vertices.clear();
  attrib->vertex_weights.clear();
  attrib->normals.clear();
  attrib->texcoords.clear();
  attrib->texcoord_ws.clear();
  attrib->colors.clear();
  attrib->skin_weights.clear();
  attrib->indices.clear();
  attrib->face_num_verts.clear();
  attrib->material_ids.clear();
  shapes->clear();
  if (materials) materials->clear();

  if (buf_len < 1) return true;  // empty buffer is not an error

  if (!buf) {
    if (err) *err = "buf must not be null when buf_len > 0.";
    return false;
  }

  // Ensure buffer ends with a newline for safe tokenization (avoids
  // one-byte over-read on the last line when parsing directly from the
  // buffer without per-line copies).
  const char *work_buf = buf;
  size_t work_len = buf_len;
  std::vector<char> buf_with_sentinel;
  if (buf[buf_len - 1] != '\n') {
    buf_with_sentinel.assign(buf, buf + buf_len);
    buf_with_sentinel.push_back('\n');
    work_buf = buf_with_sentinel.data();
    work_len = buf_with_sentinel.size();
  }

  // Determine thread count
  int num_threads = 1;
#ifdef TINYOBJLOADER_USE_MULTITHREADING
  if (config.num_threads < 0) {
    num_threads = static_cast<int>(std::thread::hardware_concurrency());
    if (num_threads < 1) num_threads = 1;
  } else if (config.num_threads > 1) {
    num_threads = config.num_threads;
  }
  if (num_threads > kOptMaxThreads) num_threads = kOptMaxThreads;
#else
#endif

  // ---- Phase 1: find line boundaries ----
  std::vector<LineInfo> all_line_infos;

#if defined(TINYOBJLOADER_USE_SIMD) && \
    (defined(TINYOBJLOADER_SIMD_SSE2) || defined(TINYOBJLOADER_SIMD_AVX2) || \
     defined(TINYOBJLOADER_SIMD_NEON))
  {
    std::vector<size_t> nl_positions;
    nl_positions.reserve(work_len / 64);
    simd_find_newlines(work_buf, work_len, nl_positions);
    simd_build_line_infos(work_buf, work_len, nl_positions, all_line_infos);
  }
#else
  {
    all_line_infos.reserve(work_len / 64);
    scalar_find_line_infos(work_buf, 0, work_len, all_line_infos);
  }
#endif

  const size_t total_lines = all_line_infos.size();
  if (total_lines == 0) return true;

  // Fast buffer-level check for legacy-only tokens (replaces per-line scan)
  if (opt_buffer_requires_legacy_fallback(work_buf, work_len)) {
    std::string input(buf, buf + buf_len);
    std::istringstream iss(input);
    attrib_t legacy_attrib;
    std::vector<shape_t> legacy_shapes;
    std::vector<material_t> legacy_materials;
    MaterialFileReader mat_reader(mtl_basedir);
    MaterialReader *reader =
        enable_mtllib_loading ? static_cast<MaterialReader *>(&mat_reader) : NULL;
    const bool ok = LoadObj(&legacy_attrib, &legacy_shapes, &legacy_materials,
                            warn, err, &iss, reader, config.triangulate, false);
    if (!ok) {
      return false;
    }
    ConvertLegacyResultToBasic(legacy_attrib, legacy_shapes, legacy_materials,
                               attrib, shapes, materials);
    return true;
  }

  // ---- Phase 2: parse lines (compact storage) ----
  //   Single-threaded or multi-threaded depending on compile option.

  const size_t num_t = static_cast<size_t>(num_threads);
  std::vector<OptThreadData> thread_data(num_t);

#ifdef TINYOBJLOADER_USE_MULTITHREADING
  // Multi-threaded path
  {
    size_t lines_per_thread = total_lines / num_t;
    std::vector<std::thread> workers;
    workers.reserve(num_t);

    for (int t = 0; t < num_threads; t++) {
      size_t start = static_cast<size_t>(t) * lines_per_thread;
      size_t end = (t == num_threads - 1)
                       ? total_lines
                       : (static_cast<size_t>(t) + 1) * lines_per_thread;

      workers.emplace_back([&, t, start, end]() {
        OptThreadData &td = thread_data[static_cast<size_t>(t)];
        size_t est_lines = end - start;
        td.v_pos.reserve(est_lines * 2);
        td.faces.reserve(est_lines / 3);
        td.seq.reserve(est_lines / 3);
        OptFloatCache *tc = nullptr;
#ifndef TINYOBJLOADER_DISABLE_FAST_FLOAT
        OptFloatCache thread_cache(config.float_cache_max_nodes,
                                   config.fp32_cache);
        if (config.float_cache) tc = &thread_cache;
#endif
        for (size_t i = start; i < end; i++) {
          opt_parseLineToThreadData(td, &work_buf[all_line_infos[i].pos],
                                     all_line_infos[i].len, config.triangulate,
                                     i + 1, tc);
          if (td.error_line != 0) break;
        }
      });
    }
    for (auto &w : workers) w.join();
  }

#else
  // Single-threaded path
  {
    OptThreadData &td = thread_data[0];
    td.v_pos.reserve(total_lines * 2);
    td.faces.reserve(total_lines / 3);
    td.seq.reserve(total_lines / 3);
    OptFloatCache *tc = nullptr;
#ifndef TINYOBJLOADER_DISABLE_FAST_FLOAT
    OptFloatCache thread_cache(config.float_cache_max_nodes,
                               config.fp32_cache);
    if (config.float_cache) tc = &thread_cache;
#endif
    for (size_t i = 0; i < total_lines; i++) {
      opt_parseLineToThreadData(td, &work_buf[all_line_infos[i].pos],
                                 all_line_infos[i].len, config.triangulate,
                                 i + 1, tc);
      if (td.error_line != 0) break;
    }
  }
#endif

  size_t first_error_line = 0;
  std::string first_error_message;
  for (size_t t = 0; t < num_t; t++) {
    if (thread_data[t].error_line == 0) continue;
    if (first_error_line == 0 || thread_data[t].error_line < first_error_line) {
      first_error_line = thread_data[t].error_line;
      first_error_message = thread_data[t].error_message;
    }
  }

  // ---- Phase 3: process sequential material state ----
  // Compute v/vn/vt prefix sums across threads
  std::vector<size_t> v_prefix(num_t), vn_prefix(num_t), vt_prefix(num_t);
  v_prefix[0] = vn_prefix[0] = vt_prefix[0] = 0;
  for (size_t t = 1; t < num_t; t++) {
    v_prefix[t] = v_prefix[t - 1] + thread_data[t - 1].num_v;
    vn_prefix[t] = vn_prefix[t - 1] + thread_data[t - 1].num_vn;
    vt_prefix[t] = vt_prefix[t - 1] + thread_data[t - 1].num_vt;
  }

  std::map<std::string, int> material_map;
  std::vector<int> initial_material_id(num_t, -1);
  size_t eof_pending_degenerate_faces = 0;
  size_t phase3_error_line = 0;
  std::string phase3_error_message;
  {
    MaterialFileReader mat_file_reader(mtl_basedir);
    std::set<std::string> material_filenames;
    std::vector<material_t> temp_materials;
    std::vector<material_t> *material_dst = materials ? materials : &temp_materials;
    int running_material_id = -1;
    size_t pending_degenerate_faces = 0;
    for (size_t t = 0; t < num_t; t++) {
      if (phase3_error_line != 0) {
        break;
      }

      initial_material_id[t] = running_material_id;
      const OptThreadData &td = thread_data[t];

      for (size_t si = 0; si < td.seq.size(); si++) {
        const OptSeqEntry &entry = td.seq[si];

        if (entry.kind == OptSeqEntry::SEQ_FACE) {
          const OptFaceCmd &face = td.faces[entry.index];

          // Compute running counts from prefix sums + per-face stored counts
          int rv = static_cast<int>(v_prefix[t] + face.v_count_before);
          int rvn = static_cast<int>(vn_prefix[t] + face.vn_count_before);
          int rvt = static_cast<int>(vt_prefix[t] + face.vt_count_before);

          if (first_error_line != 0 && face.source_line >= first_error_line) {
            continue;
          }

          if (face.degenerate) {
            pending_degenerate_faces++;
            continue;
          }

          for (size_t k = 0; k < face.face_vertex_count; k++) {
            const opt_index_t &raw = face.face_indices()[k];
            int resolved_idx = -1;

            if (!opt_validateAndResolveFaceIndexLikeLegacy(
                    raw.vertex_index, rv, false, source_name,
                    face.source_line, warn, &resolved_idx)) {
              phase3_error_line = face.source_line;
              phase3_error_message =
                  "failed to parse `f' line (invalid vertex index)";
              break;
            }

            if (raw.texcoord_index != opt_index_t::kNotPresent) {
              if (!opt_validateAndResolveFaceIndexLikeLegacy(
                      raw.texcoord_index, rvt, true, source_name,
                      face.source_line, warn, &resolved_idx)) {
                phase3_error_line = face.source_line;
                phase3_error_message =
                    "failed to parse `f' line (invalid vertex index)";
                break;
              }
            }

            if (raw.normal_index != opt_index_t::kNotPresent) {
              if (!opt_validateAndResolveFaceIndexLikeLegacy(
                      raw.normal_index, rvn, true, source_name,
                      face.source_line, warn, &resolved_idx)) {
                phase3_error_line = face.source_line;
                phase3_error_message =
                    "failed to parse `f' line (invalid vertex index)";
                break;
              }
            }
          }

          if (phase3_error_line != 0) {
            break;
          }

          continue;
        }

        // SEQ_META
        OptMetaCmd &meta = thread_data[t].metas[entry.index];

        if (first_error_line != 0 && meta.source_line >= first_error_line) {
          continue;
        }

        if (meta.type == OPT_CMD_G || meta.type == OPT_CMD_O) {
          for (size_t deg = 0; deg < pending_degenerate_faces; deg++) {
            if (warn) {
              (*warn) += "Degenerated face found\n.";
            }
          }
          pending_degenerate_faces = 0;
        }

        if (meta.type == OPT_CMD_G && meta.group_name_empty) {
          if (warn) {
            std::stringstream ss;
            ss << "Empty group name. line: " << meta.source_line << "\n";
            (*warn) += ss.str();
          }
          continue;
        }

        if (meta.type == OPT_CMD_MTLLIB) {
          if (!enable_mtllib_loading) {
            continue;
          }

          std::string line_rest;
          if (meta.str_ptr && meta.str_len > 0) {
            line_rest.assign(meta.str_ptr, meta.str_len);
          }

          std::vector<std::string> filenames;
          SplitString(line_rest, ' ', '\\', filenames);
          RemoveEmptyTokens(&filenames);

          if (filenames.empty()) {
            if (warn) {
              std::stringstream ss;
              ss << "Looks like empty filename for mtllib. Use default "
                    "material (line "
                 << meta.source_line << ".)\n";
              (*warn) += ss.str();
            }
            continue;
          }

          bool found = false;
          for (size_t s = 0; s < filenames.size(); s++) {
            if (material_filenames.count(filenames[s]) > 0) {
              found = true;
              continue;
            }

            std::string warn_mtl;
            std::string err_mtl;
            bool ok = mat_file_reader(filenames[s], material_dst, &material_map,
                                      &warn_mtl, &err_mtl);

            if (warn && !warn_mtl.empty()) {
              (*warn) += warn_mtl;
            }

            if (err && !err_mtl.empty()) {
              (*err) += err_mtl;
            }

            if (ok) {
              found = true;
              material_filenames.insert(filenames[s]);
              break;
            }
          }

          if (!found) {
            if (warn) {
              (*warn) +=
                  "Failed to load material file(s). Use default material.\n";
            }
          }
          continue;
        }

        if (meta.type == OPT_CMD_USEMTL) {
          std::string mat_name;
          if (meta.str_ptr && meta.str_len > 0) {
            mat_name.assign(meta.str_ptr, meta.str_len);
          }
          while (!mat_name.empty() &&
                 (mat_name.back() == '\r' || mat_name.back() == '\n')) {
            mat_name.pop_back();
          }

          std::map<std::string, int>::const_iterator it = material_map.find(mat_name);
          if (it != material_map.end()) {
            meta.resolved_material_id = it->second;
          } else {
            meta.resolved_material_id = -1;
            if (warn) {
              (*warn) += "material [ '" + mat_name + "' ] not found in .mtl\n";
            }
          }
          running_material_id = meta.resolved_material_id;
          continue;
        }
      }

    }

    if (first_error_line == 0) {
      eof_pending_degenerate_faces = pending_degenerate_faces;
    }
  }

  if (phase3_error_line != 0) {
    if (err) {
      std::stringstream ss;
      ss << "Failed parse line(line " << phase3_error_line << "). "
         << phase3_error_message << "\n";
      if (!err->empty()) {
        (*err) += ss.str();
      } else {
        (*err) = ss.str();
      }
    }
    return false;
  }

  if (first_error_line != 0) {
    if (err) {
      std::stringstream ss;
      ss << "Failed parse line(line " << first_error_line << "). "
         << first_error_message << "\n";
      if (!err->empty()) {
        (*err) += ss.str();
      } else {
        (*err) = ss.str();
      }
    }
    return false;
  }

  // ---- Phase 4: merge results ----
  size_t num_v = 0, num_vn = 0, num_vt = 0;
  size_t total_idx = 0;   // total index_t entries across all faces
  size_t total_faces = 0; // total emitted faces
  for (size_t t = 0; t < num_t; t++) {
    num_v += thread_data[t].num_v;
    num_vn += thread_data[t].num_vn;
    num_vt += thread_data[t].num_vt;
    total_idx += thread_data[t].num_f_indices;
    total_faces += thread_data[t].num_f_faces;
  }

  // Guard against size_t overflow in vertex/normal/texcoord allocation
  if (num_v > SIZE_MAX / 3 || num_vn > SIZE_MAX / 3 || num_vt > SIZE_MAX / 2) {
    if (err) {
      (*err) += "Integer overflow in vertex/normal/texcoord count.\n";
    }
    return false;
  }

  attrib->vertices.resize(num_v * 3);
  attrib->vertex_weights.resize(num_v, real_t(1.0));
  attrib->normals.resize(num_vn * 3);
  attrib->texcoords.resize(num_vt * 2);
  attrib->texcoord_ws.resize(num_vt, real_t(0.0));
  std::vector<unsigned int> all_smoothing_group_ids(
      static_cast<size_t>(total_faces), 0);
  attrib->indices.resize(total_idx);
  attrib->face_num_verts.resize(static_cast<size_t>(total_faces));
  attrib->material_ids.resize(static_cast<size_t>(total_faces), -1);
  std::vector<unsigned char> thread_saw_any_vertex_color(num_t, 0);
  std::vector<unsigned char> thread_missing_vertex_color(num_t, 0);
  std::vector<real_t> all_colors(num_v * 3, real_t(1.0));

  // Compute per-thread offsets
  std::vector<size_t> v_off(num_t), n_off(num_t), t_off(num_t), f_off(num_t),
      face_off(num_t);
  v_off[0] = n_off[0] = t_off[0] = f_off[0] = face_off[0] = 0;
  for (size_t t = 1; t < num_t; t++) {
    v_off[t] = v_off[t - 1] + thread_data[t - 1].num_v;
    n_off[t] = n_off[t - 1] + thread_data[t - 1].num_vn;
    t_off[t] = t_off[t - 1] + thread_data[t - 1].num_vt;
    f_off[t] = f_off[t - 1] + thread_data[t - 1].num_f_indices;
    face_off[t] = face_off[t - 1] + thread_data[t - 1].num_f_faces;
  }

  // Carry parser state that persists across lines, such as smoothing groups,
  // across thread chunk boundaries before merging in parallel.
  std::vector<unsigned int> initial_smoothing_group_id(num_t, 0);
  unsigned int running_smoothing_group_id = 0;
  for (size_t t = 0; t < num_t; t++) {
    initial_smoothing_group_id[t] = running_smoothing_group_id;
    for (size_t mi = 0; mi < thread_data[t].metas.size(); mi++) {
      if (thread_data[t].metas[mi].type == OPT_CMD_S) {
        running_smoothing_group_id = thread_data[t].metas[mi].smoothing_group_id;
      }
    }
  }

  // Merge parsed data into final arrays
  std::vector<std::vector<unsigned int> > command_written_faces(num_t);
  for (size_t t = 0; t < num_t; t++) {
    command_written_faces[t].assign(thread_data[t].faces.size(), 0);
  }
  std::vector<size_t> written_index_counts(num_t, 0);
  std::vector<size_t> written_face_counts(num_t, 0);
  std::vector<size_t> merge_error_lines(num_t, 0);
  std::vector<std::string> merge_error_messages(num_t);
  std::vector<int> thread_greatest_v_idx(num_t, -1);
  std::vector<int> thread_greatest_vn_idx(num_t, -1);
  std::vector<int> thread_greatest_vt_idx(num_t, -1);
  auto merge_vertex_thread = [&](size_t t) {
    const OptThreadData &td = thread_data[t];
    // Bulk copy positions
    if (!td.v_pos.empty()) {
      std::memcpy(&attrib->vertices[v_off[t] * 3], td.v_pos.data(),
                  td.v_pos.size() * sizeof(real_t));
    }
    // Copy weights
    if (td.saw_any_weight) {
      for (size_t i = 0; i < td.num_v; i++)
        attrib->vertex_weights[v_off[t] + i] = td.v_weight[i];
    }
    // Copy colors
    if (td.saw_any_color) {
      std::memcpy(&all_colors[v_off[t] * 3], td.v_color.data(),
                  td.v_color.size() * sizeof(real_t));
      thread_saw_any_vertex_color[t] = 1;
      if (td.saw_missing_color) thread_missing_vertex_color[t] = 1;
    } else if (td.saw_missing_color) {
      thread_missing_vertex_color[t] = 1;
    }
    // Bulk copy normals
    if (!td.vn_data.empty()) {
      std::memcpy(&attrib->normals[n_off[t] * 3], td.vn_data.data(),
                  td.vn_data.size() * sizeof(real_t));
    }
    // Bulk copy texcoords
    if (!td.vt_data.empty()) {
      std::memcpy(&attrib->texcoords[t_off[t] * 2], td.vt_data.data(),
                  td.vt_data.size() * sizeof(real_t));
    }
    if (td.saw_any_texcoord_w) {
      for (size_t i = 0; i < td.num_vt; i++)
        attrib->texcoord_ws[t_off[t] + i] = td.vt_w[i];
    }
  };
  auto merge_thread = [&](size_t t) {
    const OptThreadData &td = thread_data[t];
    size_t fc = f_off[t], fcc = face_off[t];
    int current_mat_id = initial_material_id[t];
    unsigned int current_smoothing_id = initial_smoothing_group_id[t];
    int greatest_v_idx = -1;
    int greatest_vn_idx = -1;
    int greatest_vt_idx = -1;

    for (size_t si = 0; si < td.seq.size(); si++) {
      const OptSeqEntry &entry = td.seq[si];
      if (entry.kind == OptSeqEntry::SEQ_META) {
        const OptMetaCmd &meta = td.metas[entry.index];
        if (meta.type == OPT_CMD_USEMTL)
          current_mat_id = meta.resolved_material_id;
        else if (meta.type == OPT_CMD_S)
          current_smoothing_id = meta.smoothing_group_id;
        continue;
      }
      // SEQ_FACE
      const OptFaceCmd &face = td.faces[entry.index];
      if (face.degenerate) {
        command_written_faces[t][entry.index] = 0;
        continue;
      }
      // Compute running counts for index resolution
      size_t vc = v_off[t] + face.v_count_before;
      size_t nc = n_off[t] + face.vn_count_before;
      size_t tc = t_off[t] + face.vt_count_before;

      std::vector<index_t> resolved_face(face.face_vertex_count);
      for (size_t k = 0; k < face.face_vertex_count; k++) {
        const opt_index_t &vi = face.face_indices()[k];
        index_t idx;
        if (!opt_resolveIndexLikeLegacy(vi.vertex_index,
                                        static_cast<int>(vc),
                                        &idx.vertex_index, false)) {
          merge_error_lines[t] = face.source_line;
          merge_error_messages[t] =
              "failed to parse `f' line (invalid vertex index)";
          return;
        }
        opt_updateGreatestIndex(idx.vertex_index, &greatest_v_idx);
        if (vi.texcoord_index == opt_index_t::kNotPresent) {
          idx.texcoord_index = -1;
        } else {
          if (!opt_resolveIndexLikeLegacy(vi.texcoord_index,
                                          static_cast<int>(tc),
                                          &idx.texcoord_index, true)) {
            merge_error_lines[t] = face.source_line;
            merge_error_messages[t] =
                "failed to parse `f' line (invalid vertex index)";
            return;
          }
          if (idx.texcoord_index >= 0) {
            opt_updateGreatestIndex(idx.texcoord_index, &greatest_vt_idx);
          }
        }
        if (vi.normal_index == opt_index_t::kNotPresent) {
          idx.normal_index = -1;
        } else {
          if (!opt_resolveIndexLikeLegacy(vi.normal_index,
                                          static_cast<int>(nc),
                                          &idx.normal_index, true)) {
            merge_error_lines[t] = face.source_line;
            merge_error_messages[t] =
                "failed to parse `f' line (invalid vertex index)";
            return;
          }
          if (idx.normal_index >= 0) {
            opt_updateGreatestIndex(idx.normal_index, &greatest_vn_idx);
          }
        }
        resolved_face[k] = idx;
      }
      size_t written_index_count = 0;
      size_t written_face_count = 0;
      if (config.triangulate) {
        written_index_count = opt_triangulate_face(
            attrib->vertices, resolved_face.data(), resolved_face.size(),
            &attrib->indices[fc]);
        written_face_count = written_index_count / 3;
      } else {
        written_index_count = resolved_face.size();
        written_face_count = 1;
        for (size_t k = 0; k < resolved_face.size(); k++) {
          attrib->indices[fc + k] = resolved_face[k];
        }
      }
      for (size_t k = 0; k < written_face_count; k++) {
        attrib->face_num_verts[fcc + k] = config.triangulate
                                              ? 3
                                              : static_cast<int>(face.emitted_face_verts);
        attrib->material_ids[fcc + k] = current_mat_id;
        all_smoothing_group_ids[fcc + k] = current_smoothing_id;
      }
      command_written_faces[t][entry.index] =
          static_cast<unsigned int>(written_face_count);
      fc += written_index_count;
      fcc += written_face_count;
    }
    written_index_counts[t] = fc - f_off[t];
    written_face_counts[t] = fcc - face_off[t];
    thread_greatest_v_idx[t] = greatest_v_idx;
    thread_greatest_vn_idx[t] = greatest_vn_idx;
    thread_greatest_vt_idx[t] = greatest_vt_idx;
  };

#ifdef TINYOBJLOADER_USE_MULTITHREADING
  if (num_threads > 1) {
    std::vector<std::thread> workers;
    workers.reserve(num_t);
    for (size_t t = 0; t < num_t; t++) {
      workers.emplace_back([&, t]() { merge_vertex_thread(t); });
    }
    for (auto &w : workers) w.join();
    workers.clear();
    for (size_t t = 0; t < num_t; t++) {
      workers.emplace_back([&, t]() { merge_thread(t); });
    }
    for (auto &w : workers) w.join();
  } else {
    for (size_t t = 0; t < num_t; t++) merge_vertex_thread(t);
    for (size_t t = 0; t < num_t; t++) merge_thread(t);
  }
#else
  for (size_t t = 0; t < num_t; t++) merge_vertex_thread(t);
  for (size_t t = 0; t < num_t; t++) merge_thread(t);
#endif

  size_t first_merge_error_line = 0;
  std::string first_merge_error_message;
  for (size_t t = 0; t < merge_error_lines.size(); t++) {
    if (merge_error_lines[t] == 0) continue;
    if (first_merge_error_line == 0 ||
        merge_error_lines[t] < first_merge_error_line) {
      first_merge_error_line = merge_error_lines[t];
      first_merge_error_message = merge_error_messages[t];
    }
  }

  if (first_merge_error_line != 0) {
    attrib->vertices.clear();
    attrib->vertex_weights.clear();
    attrib->normals.clear();
    attrib->texcoords.clear();
    attrib->texcoord_ws.clear();
    attrib->colors.clear();
    attrib->skin_weights.clear();
    attrib->indices.clear();
    attrib->face_num_verts.clear();
    attrib->material_ids.clear();
    shapes->clear();
    if (materials) materials->clear();

    if (err) {
      std::stringstream ss;
      ss << "Failed parse line(line " << first_merge_error_line << "). "
         << first_merge_error_message << "\n";
      if (!err->empty()) {
        (*err) += ss.str();
      } else {
        (*err) = ss.str();
      }
    }
    return false;
  }

  bool saw_any_vertex_color = false;
  bool saw_missing_vertex_color = false;
  for (size_t t = 0; t < num_t; t++) {
    saw_any_vertex_color =
        saw_any_vertex_color || (thread_saw_any_vertex_color[t] != 0);
    saw_missing_vertex_color =
        saw_missing_vertex_color || (thread_missing_vertex_color[t] != 0);
  }

  if (saw_any_vertex_color && !saw_missing_vertex_color) {
    attrib->colors.swap(all_colors);
  } else {
    attrib->colors.clear();
  }

  size_t actual_num_indices = 0;
  size_t actual_num_faces = 0;
  int greatest_v_idx = -1;
  int greatest_vn_idx = -1;
  int greatest_vt_idx = -1;
  for (size_t t = 0; t < num_t; t++) {
    actual_num_indices += written_index_counts[t];
    actual_num_faces += written_face_counts[t];
    opt_updateGreatestIndex(thread_greatest_v_idx[t], &greatest_v_idx);
    opt_updateGreatestIndex(thread_greatest_vn_idx[t], &greatest_vn_idx);
    opt_updateGreatestIndex(thread_greatest_vt_idx[t], &greatest_vt_idx);
  }

  // Compact face arrays to remove gaps left by threads that wrote fewer
  // indices/faces than pre-allocated (e.g. triangulation returning 0 for
  // faces with out-of-bounds vertex indices).
  if (num_t > 1) {
    size_t dst_idx = 0;
    size_t dst_face = 0;
    for (size_t t = 0; t < num_t; t++) {
      size_t src_idx = f_off[t];
      size_t src_face = face_off[t];
      size_t idx_count = written_index_counts[t];
      size_t fc_count = written_face_counts[t];
      if (dst_idx != src_idx && idx_count > 0) {
        std::memmove(&attrib->indices[dst_idx], &attrib->indices[src_idx],
                     idx_count * sizeof(index_t));
      }
      if (dst_face != src_face && fc_count > 0) {
        std::memmove(&attrib->face_num_verts[dst_face],
                     &attrib->face_num_verts[src_face],
                     fc_count * sizeof(int));
        std::memmove(&attrib->material_ids[dst_face],
                     &attrib->material_ids[src_face],
                     fc_count * sizeof(int));
        std::memmove(&all_smoothing_group_ids[dst_face],
                     &all_smoothing_group_ids[src_face],
                     fc_count * sizeof(unsigned int));
      }
      dst_idx += idx_count;
      dst_face += fc_count;
    }
  }

  attrib->indices.resize(actual_num_indices);
  attrib->face_num_verts.resize(actual_num_faces);
  attrib->material_ids.resize(actual_num_faces);
  all_smoothing_group_ids.resize(actual_num_faces);

  opt_appendOutOfBoundsWarnings(
      warn, greatest_v_idx, greatest_vn_idx, greatest_vt_idx,
      static_cast<int>(attrib->vertices.size() / 3),
      static_cast<int>(attrib->normals.size() / 3),
      static_cast<int>(attrib->texcoords.size() / 2),
      all_line_infos.size() + 1);
  for (size_t deg = 0; deg < eof_pending_degenerate_faces; deg++) {
    if (warn) {
      (*warn) += "Degenerated face found\n.";
    }
  }

  // ---- Phase 5: construct shapes ----
  {
    // Precompute prefix-sum of index offsets for O(1) slicing
    const size_t total_faces = attrib->face_num_verts.size();
    std::vector<size_t> idx_prefix(total_faces + 1);
    idx_prefix[0] = 0;
    for (size_t fi = 0; fi < total_faces; fi++) {
      idx_prefix[fi + 1] =
          idx_prefix[fi] + static_cast<size_t>(attrib->face_num_verts[fi]);
    }

    size_t face_count = 0;
    basic_shape_t<> shape;
    size_t face_prev_offset = 0;
    bool shape_has_face_record = false;
    bool have_active_shape_name = false;

    for (size_t t = 0; t < num_t; t++) {
      const OptThreadData &td = thread_data[t];
      for (size_t si = 0; si < td.seq.size(); si++) {
        const OptSeqEntry &entry = td.seq[si];
        if (entry.kind == OptSeqEntry::SEQ_META) {
          const OptMetaCmd &meta = td.metas[entry.index];
          if (meta.type == OPT_CMD_O || meta.type == OPT_CMD_G) {
            std::string name;
            if (meta.type == OPT_CMD_O && meta.str_ptr) {
              name.assign(meta.str_ptr, meta.str_len);
            } else if (!meta.str_storage.empty()) {
              name = meta.str_storage;
            } else if (meta.str_ptr) {
              name.assign(meta.str_ptr, meta.str_len);
            }
            while (!name.empty() &&
                   (name.back() == '\r' || name.back() == '\n'))
              name.pop_back();

            if (face_count == 0) {
              shape.name = name;
              face_prev_offset = 0;
              have_active_shape_name = true;
            } else {
              if (!have_active_shape_name) {
                // faces before first group/object
                basic_shape_t<> prev_shape;
                prev_shape.mesh.num_face_vertices.assign(
                    attrib->face_num_verts.begin(),
                    attrib->face_num_verts.begin() +
                        static_cast<std::ptrdiff_t>(face_count));
                prev_shape.mesh.indices.assign(
                    attrib->indices.begin(),
                    attrib->indices.begin() +
                        static_cast<std::ptrdiff_t>(idx_prefix[face_count]));
                prev_shape.mesh.material_ids.assign(
                    attrib->material_ids.begin(),
                    attrib->material_ids.begin() +
                        static_cast<std::ptrdiff_t>(face_count));
                prev_shape.mesh.smoothing_group_ids.assign(
                    all_smoothing_group_ids.begin(),
                    all_smoothing_group_ids.begin() +
                        static_cast<std::ptrdiff_t>(face_count));
                shapes->push_back(std::move(prev_shape));
              } else if (face_count > face_prev_offset) {
                // push previous shape
                basic_shape_t<> prev_shape;
                prev_shape.name = shape.name;
                prev_shape.mesh.num_face_vertices.assign(
                    attrib->face_num_verts.begin() +
                        static_cast<std::ptrdiff_t>(face_prev_offset),
                    attrib->face_num_verts.begin() +
                        static_cast<std::ptrdiff_t>(face_count));
                prev_shape.mesh.indices.assign(
                    attrib->indices.begin() +
                        static_cast<std::ptrdiff_t>(idx_prefix[face_prev_offset]),
                    attrib->indices.begin() +
                        static_cast<std::ptrdiff_t>(idx_prefix[face_count]));
                prev_shape.mesh.material_ids.assign(
                    attrib->material_ids.begin() +
                        static_cast<std::ptrdiff_t>(face_prev_offset),
                    attrib->material_ids.begin() +
                        static_cast<std::ptrdiff_t>(face_count));
                prev_shape.mesh.smoothing_group_ids.assign(
                    all_smoothing_group_ids.begin() +
                        static_cast<std::ptrdiff_t>(face_prev_offset),
                    all_smoothing_group_ids.begin() +
                        static_cast<std::ptrdiff_t>(face_count));
                shapes->push_back(std::move(prev_shape));
              }
              shape.name = name;
              face_prev_offset = face_count;
              have_active_shape_name = true;
            }
            shape_has_face_record = false;
          }
        } else {
          // SEQ_FACE
          shape_has_face_record = true;
          face_count += command_written_faces[t][entry.index];
        }
      }
    }

    // Final shape
    if (face_count > face_prev_offset || shape_has_face_record) {
      basic_shape_t<> final_shape;
      final_shape.name = shape.name;
      final_shape.mesh.num_face_vertices.assign(
          attrib->face_num_verts.begin() +
              static_cast<std::ptrdiff_t>(face_prev_offset),
          attrib->face_num_verts.begin() +
              static_cast<std::ptrdiff_t>(face_count));
      final_shape.mesh.indices.assign(
          attrib->indices.begin() +
              static_cast<std::ptrdiff_t>(idx_prefix[face_prev_offset]),
          attrib->indices.begin() +
              static_cast<std::ptrdiff_t>(idx_prefix[face_count]));
      final_shape.mesh.material_ids.assign(
          attrib->material_ids.begin() +
              static_cast<std::ptrdiff_t>(face_prev_offset),
          attrib->material_ids.begin() +
              static_cast<std::ptrdiff_t>(face_count));
      final_shape.mesh.smoothing_group_ids.assign(
          all_smoothing_group_ids.begin() +
              static_cast<std::ptrdiff_t>(face_prev_offset),
          all_smoothing_group_ids.begin() +
              static_cast<std::ptrdiff_t>(face_count));
      shapes->push_back(std::move(final_shape));
    }
  }

  return true;
}

// ---- LoadObjOpt (buffer version, public API) ----

bool LoadObjOpt(basic_attrib_t<> *attrib,
                std::vector<basic_shape_t<>> *shapes,
                std::vector<material_t> *materials,
                std::string *warn, std::string *err,
                const char *buf, size_t buf_len,
                const OptLoadConfig &config) {
  return LoadObjOpt_internal(attrib, shapes, materials, warn, err,
                             buf, buf_len, std::string(), "<stream>", false,
                             config);
}

// ---- LoadObjOpt (file version) ----

bool LoadObjOpt(basic_attrib_t<> *attrib,
                std::vector<basic_shape_t<>> *shapes,
                std::vector<material_t> *materials,
                std::string *warn, std::string *err,
                const char *filename,
                const char *mtl_basedir,
                const OptLoadConfig &config) {
  if (!filename) {
    if (err) *err = "filename is null.";
    return false;
  }

  std::string filepath(filename);

  // Resolve material base directory
  std::string baseDir;
  if (mtl_basedir) {
    baseDir = mtl_basedir;
  } else {
    // Extract directory from filename
    size_t pos = filepath.find_last_of("/\\");
    if (pos != std::string::npos) {
      baseDir = filepath.substr(0, pos + 1);
    }
  }
  if (!baseDir.empty()) {
#ifndef _WIN32
    const char dirsep = '/';
#else
    const char dirsep = '\\';
#endif
    if (baseDir[baseDir.length() - 1] != dirsep) baseDir += dirsep;
  }

#ifdef TINYOBJLOADER_USE_MMAP
  {
    MappedFile mf;
    if (mf.open(filepath.c_str())) {
      return LoadObjOpt_internal(attrib, shapes, materials, warn, err,
                                 mf.data, mf.size, baseDir, filepath, true,
                                 config);
    }
  }
#endif

#ifdef _WIN32
  std::ifstream ifs(LongPathW(UTF8ToWchar(filepath)).c_str(),
                    std::ios::binary | std::ios::ate);
#else
  std::ifstream ifs(filepath.c_str(), std::ios::binary | std::ios::ate);
#endif
  if (!ifs.is_open()) {
    if (err) *err = "Cannot open file: " + filepath;
    return false;
  }

  std::streamsize fsize = ifs.tellg();
  ifs.seekg(0, std::ios::beg);

  if (fsize <= 0) {
    return LoadObjOpt_internal(attrib, shapes, materials, warn, err,
                               "", static_cast<size_t>(0), baseDir, filepath,
                               true, config);
  }

  std::vector<char> buf(static_cast<size_t>(fsize));
  if (!ifs.read(buf.data(), fsize)) {
    if (err) *err = "Failed to read file: " + filepath;
    return false;
  }
  ifs.close();

  // Parse the in-memory file buffer with baseDir for mtllib resolution.
  return LoadObjOpt_internal(attrib, shapes, materials, warn, err,
                             buf.data(), static_cast<size_t>(fsize), baseDir,
                             filepath, true, config);
}

// ---- LoadObjOptTyped internals ----

static bool LoadObjOptTyped_internal(OptResult *result,
                                     std::string *warn, std::string *err,
                                     const char *buf, size_t buf_len,
                                     const std::string &mtl_basedir,
                                     const std::string &source_name,
                                     bool enable_mtllib_loading,
                                     const OptLoadConfig &config) {
  using namespace opt_internal;

  ArenaAllocator &arena = result->arena;
  OptAttrib &attrib = result->attrib;
  result->valid = false;

  if (buf_len < 1) {
    result->valid = true;
    return true;
  }
  if (!buf) {
    if (err) *err = "buf must not be null when buf_len > 0.";
    return false;
  }

  const char *work_buf = buf;
  size_t work_len = buf_len;
  std::vector<char> buf_with_sentinel;
  if (buf[buf_len - 1] != '\n') {
    buf_with_sentinel.assign(buf, buf + buf_len);
    buf_with_sentinel.push_back('\n');
    work_buf = buf_with_sentinel.data();
    work_len = buf_with_sentinel.size();
  }

  int num_threads = 1;
#ifdef TINYOBJLOADER_USE_MULTITHREADING
  if (config.num_threads < 0) {
    num_threads = static_cast<int>(std::thread::hardware_concurrency());
    if (num_threads < 1) num_threads = 1;
  } else if (config.num_threads > 1) {
    num_threads = config.num_threads;
  }
  if (num_threads > kOptMaxThreads) num_threads = kOptMaxThreads;
#else
#endif

  // ---- Phase 1: find line boundaries ----
  std::vector<LineInfo> all_line_infos;

#if defined(TINYOBJLOADER_USE_SIMD) && \
    (defined(TINYOBJLOADER_SIMD_SSE2) || defined(TINYOBJLOADER_SIMD_AVX2) || \
     defined(TINYOBJLOADER_SIMD_NEON))
  {
    std::vector<size_t> nl_positions;
    nl_positions.reserve(work_len / 64);
    simd_find_newlines(work_buf, work_len, nl_positions);
    simd_build_line_infos(work_buf, work_len, nl_positions, all_line_infos);
  }
#else
  {
    all_line_infos.reserve(work_len / 64);
    scalar_find_line_infos(work_buf, 0, work_len, all_line_infos);
  }
#endif

  const size_t total_lines = all_line_infos.size();
  if (total_lines == 0) {
    result->valid = true;
    return true;
  }

  // Fast buffer-level check for legacy-only tokens (replaces per-line scan)
  if (opt_buffer_requires_legacy_fallback(work_buf, work_len)) {
    basic_attrib_t<> tmp_attrib;
    std::vector<basic_shape_t<>> tmp_shapes;
    std::vector<material_t> tmp_materials;
    std::string input(buf, buf + buf_len);
    std::istringstream iss(input);
    attrib_t legacy_attrib;
    std::vector<shape_t> legacy_shapes;
    std::vector<material_t> legacy_materials;
    MaterialFileReader mat_reader(mtl_basedir);
    MaterialReader *reader =
        enable_mtllib_loading ? static_cast<MaterialReader *>(&mat_reader) : NULL;
    const bool ok = LoadObj(&legacy_attrib, &legacy_shapes, &legacy_materials,
                            warn, err, &iss, reader, config.triangulate, false);
    if (!ok) return false;
    ConvertLegacyResultToBasic(legacy_attrib, legacy_shapes, legacy_materials,
                               &tmp_attrib, &tmp_shapes, &tmp_materials);
#define TINYOBJ_COPY_VEC_TO_ARENA_(dst, src)        \
    do {                                             \
      if (!(src).empty()) {                          \
        (dst).allocate(arena, (src).size());          \
        std::memcpy((dst).data(), (src).data(),       \
                    (src).size() * sizeof((src)[0])); \
      }                                              \
    } while (0)
    TINYOBJ_COPY_VEC_TO_ARENA_(attrib.vertices, tmp_attrib.vertices);
    TINYOBJ_COPY_VEC_TO_ARENA_(attrib.vertex_weights, tmp_attrib.vertex_weights);
    TINYOBJ_COPY_VEC_TO_ARENA_(attrib.normals, tmp_attrib.normals);
    TINYOBJ_COPY_VEC_TO_ARENA_(attrib.texcoords, tmp_attrib.texcoords);
    TINYOBJ_COPY_VEC_TO_ARENA_(attrib.texcoord_ws, tmp_attrib.texcoord_ws);
    TINYOBJ_COPY_VEC_TO_ARENA_(attrib.colors, tmp_attrib.colors);
    TINYOBJ_COPY_VEC_TO_ARENA_(attrib.indices, tmp_attrib.indices);
    TINYOBJ_COPY_VEC_TO_ARENA_(attrib.face_num_verts, tmp_attrib.face_num_verts);
    TINYOBJ_COPY_VEC_TO_ARENA_(attrib.material_ids, tmp_attrib.material_ids);
#undef TINYOBJ_COPY_VEC_TO_ARENA_
    // Flatten per-shape smoothing_group_ids into attrib
    {
      size_t total_sg = 0;
      for (size_t si = 0; si < tmp_shapes.size(); si++)
        total_sg += tmp_shapes[si].mesh.smoothing_group_ids.size();
      if (total_sg > 0) {
        attrib.smoothing_group_ids.allocate(arena, total_sg);
        size_t off = 0;
        for (size_t si = 0; si < tmp_shapes.size(); si++) {
          const auto &sg = tmp_shapes[si].mesh.smoothing_group_ids;
          if (!sg.empty()) {
            std::memcpy(&attrib.smoothing_group_ids[off], sg.data(),
                        sg.size() * sizeof(sg[0]));
            off += sg.size();
          }
        }
      }
    }
    result->shapes.clear();
    size_t idx_off = 0, face_off = 0;
    for (size_t si = 0; si < tmp_shapes.size(); si++) {
      OptShapeRange sr;
      sr.name = tmp_shapes[si].name;
      sr.face_offset = face_off;
      sr.face_count = tmp_shapes[si].mesh.num_face_vertices.size();
      sr.index_offset = idx_off;
      sr.index_count = tmp_shapes[si].mesh.indices.size();
      face_off += sr.face_count;
      idx_off += sr.index_count;
      result->shapes.push_back(std::move(sr));
    }
    result->materials = tmp_materials;
    result->valid = true;
    return true;
  }

  // ---- Phase 2: parse lines (compact storage) ----
  const size_t num_t = static_cast<size_t>(num_threads);
  std::vector<OptThreadData> thread_data(num_t);

#ifdef TINYOBJLOADER_USE_MULTITHREADING
  {
    size_t lines_per_thread = total_lines / num_t;
    std::vector<std::thread> workers;
    workers.reserve(num_t);

    for (int t = 0; t < num_threads; t++) {
      size_t start = static_cast<size_t>(t) * lines_per_thread;
      size_t end = (t == num_threads - 1)
                       ? total_lines
                       : (static_cast<size_t>(t) + 1) * lines_per_thread;

      workers.emplace_back([&, t, start, end]() {
        OptThreadData &td = thread_data[static_cast<size_t>(t)];
        size_t est_lines = end - start;
        td.v_pos.reserve(est_lines * 2);
        td.faces.reserve(est_lines / 3);
        td.seq.reserve(est_lines / 3);
        OptFloatCache *tc = nullptr;
#ifndef TINYOBJLOADER_DISABLE_FAST_FLOAT
        OptFloatCache thread_cache(config.float_cache_max_nodes,
                                   config.fp32_cache);
        if (config.float_cache) tc = &thread_cache;
#endif
        for (size_t i = start; i < end; i++) {
          opt_parseLineToThreadData(td, &work_buf[all_line_infos[i].pos],
                                     all_line_infos[i].len, config.triangulate,
                                     i + 1, tc);
          if (td.error_line != 0) break;
        }
      });
    }
    for (auto &w : workers) w.join();
  }
#else
  {
    OptThreadData &td = thread_data[0];
    td.v_pos.reserve(total_lines * 2);
    td.faces.reserve(total_lines / 3);
    td.seq.reserve(total_lines / 3);
    OptFloatCache *tc = nullptr;
#ifndef TINYOBJLOADER_DISABLE_FAST_FLOAT
    OptFloatCache thread_cache(config.float_cache_max_nodes,
                               config.fp32_cache);
    if (config.float_cache) tc = &thread_cache;
#endif
    for (size_t i = 0; i < total_lines; i++) {
      opt_parseLineToThreadData(td, &work_buf[all_line_infos[i].pos],
                                 all_line_infos[i].len, config.triangulate,
                                 i + 1, tc);
      if (td.error_line != 0) break;
    }
  }
#endif

  // Check for parse errors
  size_t first_error_line = 0;
  std::string first_error_message;
  for (size_t t = 0; t < num_t; t++) {
    if (thread_data[t].error_line == 0) continue;
    if (first_error_line == 0 || thread_data[t].error_line < first_error_line) {
      first_error_line = thread_data[t].error_line;
      first_error_message = thread_data[t].error_message;
    }
  }

  // ---- Phase 3: process sequential material state ----
  // Compute v/vn/vt prefix sums across threads
  std::vector<size_t> v_prefix(num_t), vn_prefix(num_t), vt_prefix(num_t);
  v_prefix[0] = vn_prefix[0] = vt_prefix[0] = 0;
  for (size_t t = 1; t < num_t; t++) {
    v_prefix[t] = v_prefix[t - 1] + thread_data[t - 1].num_v;
    vn_prefix[t] = vn_prefix[t - 1] + thread_data[t - 1].num_vn;
    vt_prefix[t] = vt_prefix[t - 1] + thread_data[t - 1].num_vt;
  }

  std::map<std::string, int> material_map;
  std::vector<int> initial_material_id(num_t, -1);
  size_t eof_pending_degenerate_faces = 0;
  size_t phase3_error_line = 0;
  std::string phase3_error_message;
  {
    MaterialFileReader mat_file_reader(mtl_basedir);
    std::set<std::string> material_filenames;
    std::vector<material_t> *material_dst = &result->materials;
    int running_material_id = -1;
    size_t pending_degenerate_faces = 0;

    for (size_t t = 0; t < num_t; t++) {
      if (phase3_error_line != 0) break;
      initial_material_id[t] = running_material_id;
      const OptThreadData &td = thread_data[t];

      for (size_t si = 0; si < td.seq.size(); si++) {
        const OptSeqEntry &entry = td.seq[si];

        if (entry.kind == OptSeqEntry::SEQ_FACE) {
          const OptFaceCmd &face = td.faces[entry.index];
          int rv = static_cast<int>(v_prefix[t] + face.v_count_before);
          int rvn = static_cast<int>(vn_prefix[t] + face.vn_count_before);
          int rvt = static_cast<int>(vt_prefix[t] + face.vt_count_before);

          if (first_error_line != 0 && face.source_line >= first_error_line) continue;

          if (face.degenerate) {
            pending_degenerate_faces++;
            continue;
          }

          for (size_t k = 0; k < face.face_vertex_count; k++) {
            const opt_index_t &raw = face.face_indices()[k];
            int resolved_idx = -1;
            if (!opt_validateAndResolveFaceIndexLikeLegacy(
                    raw.vertex_index, rv, false, source_name,
                    face.source_line, warn, &resolved_idx)) {
              phase3_error_line = face.source_line;
              phase3_error_message = "failed to parse `f' line (invalid vertex index)";
              break;
            }
            if (raw.texcoord_index != opt_index_t::kNotPresent) {
              if (!opt_validateAndResolveFaceIndexLikeLegacy(
                      raw.texcoord_index, rvt, true, source_name,
                      face.source_line, warn, &resolved_idx)) {
                phase3_error_line = face.source_line;
                phase3_error_message = "failed to parse `f' line (invalid vertex index)";
                break;
              }
            }
            if (raw.normal_index != opt_index_t::kNotPresent) {
              if (!opt_validateAndResolveFaceIndexLikeLegacy(
                      raw.normal_index, rvn, true, source_name,
                      face.source_line, warn, &resolved_idx)) {
                phase3_error_line = face.source_line;
                phase3_error_message = "failed to parse `f' line (invalid vertex index)";
                break;
              }
            }
          }
          if (phase3_error_line != 0) break;
          continue;
        }

        // SEQ_META
        OptMetaCmd &meta = thread_data[t].metas[entry.index];

        if (first_error_line != 0 && meta.source_line >= first_error_line) continue;

        if (meta.type == OPT_CMD_G || meta.type == OPT_CMD_O) {
          for (size_t deg = 0; deg < pending_degenerate_faces; deg++) {
            if (warn) (*warn) += "Degenerated face found\n.";
          }
          pending_degenerate_faces = 0;
        }

        if (meta.type == OPT_CMD_G && meta.group_name_empty) {
          if (warn) {
            (*warn) += "Empty group name. line: " +
                       std::to_string(meta.source_line) + "\n";
          }
          continue;
        }

        if (meta.type == OPT_CMD_MTLLIB) {
          if (!enable_mtllib_loading) continue;
          std::string line_rest;
          if (meta.str_ptr && meta.str_len > 0) {
            line_rest.assign(meta.str_ptr, meta.str_len);
          }
          std::vector<std::string> filenames;
          SplitString(line_rest, ' ', '\\', filenames);
          RemoveEmptyTokens(&filenames);
          if (filenames.empty()) {
            if (warn) {
              (*warn) += "Looks like empty filename for mtllib. Use default "
                         "material (line " +
                         std::to_string(meta.source_line) + ".)\n";
            }
            continue;
          }
          bool found = false;
          for (size_t s = 0; s < filenames.size(); s++) {
            if (material_filenames.count(filenames[s]) > 0) {
              found = true;
              continue;
            }
            std::string warn_mtl, err_mtl;
            bool ok = mat_file_reader(filenames[s], material_dst, &material_map,
                                      &warn_mtl, &err_mtl);
            if (warn && !warn_mtl.empty()) (*warn) += warn_mtl;
            if (err && !err_mtl.empty()) (*err) += err_mtl;
            if (ok) {
              found = true;
              material_filenames.insert(filenames[s]);
              break;
            }
          }
          if (!found && warn) {
            (*warn) += "Failed to load material file(s). Use default material.\n";
          }
          continue;
        }

        if (meta.type == OPT_CMD_USEMTL) {
          std::string mat_name;
          if (meta.str_ptr && meta.str_len > 0) {
            mat_name.assign(meta.str_ptr, meta.str_len);
          }
          while (!mat_name.empty() &&
                 (mat_name.back() == '\r' || mat_name.back() == '\n')) {
            mat_name.pop_back();
          }
          std::map<std::string, int>::const_iterator it = material_map.find(mat_name);
          if (it != material_map.end()) {
            meta.resolved_material_id = it->second;
          } else {
            meta.resolved_material_id = -1;
            if (warn) (*warn) += "material [ '" + mat_name + "' ] not found in .mtl\n";
          }
          running_material_id = meta.resolved_material_id;
          continue;
        }
      }

    }
    if (first_error_line == 0) {
      eof_pending_degenerate_faces = pending_degenerate_faces;
    }
  }

  if (phase3_error_line != 0) {
    if (err) {
      (*err) += "Failed parse line(line " +
                std::to_string(phase3_error_line) + "). " +
                phase3_error_message + "\n";
    }
    return false;
  }

  if (first_error_line != 0) {
    if (err) {
      (*err) += "Failed parse line(line " +
                std::to_string(first_error_line) + "). " +
                first_error_message + "\n";
    }
    return false;
  }

  // ---- Phase 4: allocate arena arrays and merge ----
  size_t num_v = 0, num_vn = 0, num_vt = 0;
  size_t total_idx = 0;   // total index_t entries across all faces
  size_t total_faces = 0; // total emitted faces
  for (size_t t = 0; t < num_t; t++) {
    num_v += thread_data[t].num_v;
    num_vn += thread_data[t].num_vn;
    num_vt += thread_data[t].num_vt;
    total_idx += thread_data[t].num_f_indices;
    total_faces += thread_data[t].num_f_faces;
  }

  // Guard against size_t overflow in vertex/normal/texcoord allocation
  if (num_v > SIZE_MAX / 3 || num_vn > SIZE_MAX / 3 || num_vt > SIZE_MAX / 2) {
    if (err) {
      (*err) += "Integer overflow in vertex/normal/texcoord count.\n";
    }
    return false;
  }

  // Determine which optional arrays are needed
  bool any_color = false;
  bool any_weight = false, any_texcoord_w = false, any_smoothing = false;
  for (size_t t = 0; t < num_t; t++) {
    if (thread_data[t].saw_any_color) any_color = true;
    if (thread_data[t].saw_any_weight) any_weight = true;
    if (thread_data[t].saw_any_texcoord_w) any_texcoord_w = true;
    if (thread_data[t].saw_any_smoothing) any_smoothing = true;
  }

  // Allocate from arena
  attrib.vertices.allocate(arena, num_v * 3);
  if (any_weight) {
    attrib.vertex_weights.allocate(arena, num_v);
    for (size_t i = 0; i < num_v; i++)
      attrib.vertex_weights[i] = real_t(1.0);
  }
  attrib.normals.allocate(arena, num_vn * 3);
  attrib.texcoords.allocate(arena, num_vt * 2);
  if (any_texcoord_w) {
    attrib.texcoord_ws.allocate(arena, num_vt);
  }
  attrib.indices.allocate(arena, total_idx);
  attrib.face_num_verts.allocate(arena, total_faces);
  attrib.material_ids.allocate(arena, total_faces);
  for (size_t i = 0; i < total_faces; i++)
    attrib.material_ids[i] = -1;

  // Smoothing group ids — only if any smoothing commands seen
  TypedArray<unsigned int> all_smoothing_group_ids;
  if (any_smoothing) {
    all_smoothing_group_ids.allocate(arena, total_faces);
  }

  // Colors — allocate temp only if any vertex has color
  TypedArray<real_t> all_colors;
  if (any_color) {
    all_colors.allocate(arena, num_v * 3);
    for (size_t i = 0; i < num_v * 3; i++)
      all_colors[i] = real_t(1.0);
  }

  // Compute per-thread offsets
  std::vector<size_t> v_off(num_t), n_off(num_t), t_off(num_t), f_off(num_t),
      face_off(num_t);
  v_off[0] = n_off[0] = t_off[0] = f_off[0] = face_off[0] = 0;
  for (size_t t = 1; t < num_t; t++) {
    v_off[t] = v_off[t - 1] + thread_data[t - 1].num_v;
    n_off[t] = n_off[t - 1] + thread_data[t - 1].num_vn;
    t_off[t] = t_off[t - 1] + thread_data[t - 1].num_vt;
    f_off[t] = f_off[t - 1] + thread_data[t - 1].num_f_indices;
    face_off[t] = face_off[t - 1] + thread_data[t - 1].num_f_faces;
  }

  // Carry smoothing group state across thread boundaries
  std::vector<unsigned int> initial_smoothing_group_id(num_t, 0);
  unsigned int running_smoothing_group_id = 0;
  for (size_t t = 0; t < num_t; t++) {
    initial_smoothing_group_id[t] = running_smoothing_group_id;
    for (size_t mi = 0; mi < thread_data[t].metas.size(); mi++) {
      if (thread_data[t].metas[mi].type == OPT_CMD_S) {
        running_smoothing_group_id = thread_data[t].metas[mi].smoothing_group_id;
      }
    }
  }

  // Per-thread face tracking for shape construction
  std::vector<std::vector<unsigned int>> command_written_faces(num_t);
  for (size_t t = 0; t < num_t; t++) {
    command_written_faces[t].assign(thread_data[t].faces.size(), 0);
  }
  std::vector<size_t> written_index_counts(num_t, 0);
  std::vector<size_t> written_face_counts(num_t, 0);
  std::vector<size_t> merge_error_lines(num_t, 0);
  std::vector<std::string> merge_error_messages(num_t);
  std::vector<int> thread_greatest_v_idx(num_t, -1);
  std::vector<int> thread_greatest_vn_idx(num_t, -1);
  std::vector<int> thread_greatest_vt_idx(num_t, -1);
  std::vector<unsigned char> thread_saw_any_vertex_color(num_t, 0);
  std::vector<unsigned char> thread_missing_vertex_color(num_t, 0);

  auto merge_vertex_thread = [&](size_t t) {
    const OptThreadData &td = thread_data[t];
    // Bulk copy positions
    if (!td.v_pos.empty()) {
      std::memcpy(&attrib.vertices[v_off[t] * 3], td.v_pos.data(),
                  td.v_pos.size() * sizeof(real_t));
    }
    // Copy weights
    if (any_weight && td.saw_any_weight) {
      for (size_t i = 0; i < td.num_v; i++)
        attrib.vertex_weights[v_off[t] + i] = td.v_weight[i];
    }
    // Copy colors
    if (any_color) {
      if (td.saw_any_color) {
        std::memcpy(&all_colors[v_off[t] * 3], td.v_color.data(),
                    td.v_color.size() * sizeof(real_t));
        thread_saw_any_vertex_color[t] = 1;
        if (td.saw_missing_color) thread_missing_vertex_color[t] = 1;
      } else if (td.saw_missing_color) {
        thread_missing_vertex_color[t] = 1;
      }
    }
    // Bulk copy normals
    if (!td.vn_data.empty()) {
      std::memcpy(&attrib.normals[n_off[t] * 3], td.vn_data.data(),
                  td.vn_data.size() * sizeof(real_t));
    }
    // Bulk copy texcoords
    if (!td.vt_data.empty()) {
      std::memcpy(&attrib.texcoords[t_off[t] * 2], td.vt_data.data(),
                  td.vt_data.size() * sizeof(real_t));
    }
    if (any_texcoord_w && td.saw_any_texcoord_w) {
      for (size_t i = 0; i < td.num_vt; i++)
        attrib.texcoord_ws[t_off[t] + i] = td.vt_w[i];
    }
  };

  auto merge_thread = [&](size_t t) {
    const OptThreadData &td = thread_data[t];
    size_t fc = f_off[t], fcc = face_off[t];
    int current_mat_id = initial_material_id[t];
    unsigned int current_smoothing_id = initial_smoothing_group_id[t];
    int greatest_v_idx = -1;
    int greatest_vn_idx = -1;
    int greatest_vt_idx = -1;

    // Reusable scratch buffer for resolved face indices
    index_t resolved_inline[8];
    std::vector<index_t> resolved_heap;

    for (size_t si = 0; si < td.seq.size(); si++) {
      const OptSeqEntry &entry = td.seq[si];
      if (entry.kind == OptSeqEntry::SEQ_META) {
        const OptMetaCmd &meta = td.metas[entry.index];
        if (meta.type == OPT_CMD_USEMTL)
          current_mat_id = meta.resolved_material_id;
        else if (meta.type == OPT_CMD_S)
          current_smoothing_id = meta.smoothing_group_id;
        continue;
      }
      // SEQ_FACE
      const OptFaceCmd &face = td.faces[entry.index];
      if (face.degenerate) {
        command_written_faces[t][entry.index] = 0;
        continue;
      }
      // Compute running counts for index resolution
      size_t vc = v_off[t] + face.v_count_before;
      size_t nc = n_off[t] + face.vn_count_before;
      size_t tc = t_off[t] + face.vt_count_before;

      // Use stack buffer for typical faces, heap only for large polygons
      index_t *resolved_face;
      if (face.face_vertex_count <= 8) {
        resolved_face = resolved_inline;
      } else {
        resolved_heap.resize(face.face_vertex_count);
        resolved_face = resolved_heap.data();
      }
      for (size_t k = 0; k < face.face_vertex_count; k++) {
        const opt_index_t &vi = face.face_indices()[k];
        index_t idx;
        if (!opt_resolveIndexLikeLegacy(vi.vertex_index,
                                        static_cast<int>(vc),
                                        &idx.vertex_index, false)) {
          merge_error_lines[t] = face.source_line;
          merge_error_messages[t] =
              "failed to parse `f' line (invalid vertex index)";
          return;
        }
        opt_updateGreatestIndex(idx.vertex_index, &greatest_v_idx);
        if (vi.texcoord_index == opt_index_t::kNotPresent) {
          idx.texcoord_index = -1;
        } else {
          if (!opt_resolveIndexLikeLegacy(vi.texcoord_index,
                                          static_cast<int>(tc),
                                          &idx.texcoord_index, true)) {
            merge_error_lines[t] = face.source_line;
            merge_error_messages[t] =
                "failed to parse `f' line (invalid vertex index)";
            return;
          }
          if (idx.texcoord_index >= 0)
            opt_updateGreatestIndex(idx.texcoord_index, &greatest_vt_idx);
        }
        if (vi.normal_index == opt_index_t::kNotPresent) {
          idx.normal_index = -1;
        } else {
          if (!opt_resolveIndexLikeLegacy(vi.normal_index,
                                          static_cast<int>(nc),
                                          &idx.normal_index, true)) {
            merge_error_lines[t] = face.source_line;
            merge_error_messages[t] =
                "failed to parse `f' line (invalid vertex index)";
            return;
          }
          if (idx.normal_index >= 0)
            opt_updateGreatestIndex(idx.normal_index, &greatest_vn_idx);
        }
        resolved_face[k] = idx;
      }
      size_t written_index_count = 0;
      size_t written_face_count = 0;
      if (config.triangulate) {
        written_index_count = opt_triangulate_face(
            attrib.vertices.data(), attrib.vertices.size(),
            resolved_face, face.face_vertex_count,
            &attrib.indices[fc]);
        written_face_count = written_index_count / 3;
      } else {
        written_index_count = face.face_vertex_count;
        written_face_count = 1;
        for (size_t k = 0; k < face.face_vertex_count; k++) {
          attrib.indices[fc + k] = resolved_face[k];
        }
      }
      for (size_t k = 0; k < written_face_count; k++) {
        attrib.face_num_verts[fcc + k] = config.triangulate
                                            ? 3
                                            : static_cast<int>(face.emitted_face_verts);
        attrib.material_ids[fcc + k] = current_mat_id;
        if (any_smoothing) {
          all_smoothing_group_ids[fcc + k] = current_smoothing_id;
        }
      }
      command_written_faces[t][entry.index] =
          static_cast<unsigned int>(written_face_count);
      fc += written_index_count;
      fcc += written_face_count;
    }
    written_index_counts[t] = fc - f_off[t];
    written_face_counts[t] = fcc - face_off[t];
    thread_greatest_v_idx[t] = greatest_v_idx;
    thread_greatest_vn_idx[t] = greatest_vn_idx;
    thread_greatest_vt_idx[t] = greatest_vt_idx;
  };

#ifdef TINYOBJLOADER_USE_MULTITHREADING
  if (num_threads > 1) {
    std::vector<std::thread> workers;
    workers.reserve(num_t);
    for (size_t t = 0; t < num_t; t++) {
      workers.emplace_back([&, t]() { merge_vertex_thread(t); });
    }
    for (auto &w : workers) w.join();
    workers.clear();
    for (size_t t = 0; t < num_t; t++) {
      workers.emplace_back([&, t]() { merge_thread(t); });
    }
    for (auto &w : workers) w.join();
  } else {
    for (size_t t = 0; t < num_t; t++) merge_vertex_thread(t);
    for (size_t t = 0; t < num_t; t++) merge_thread(t);
  }
#else
  for (size_t t = 0; t < num_t; t++) merge_vertex_thread(t);
  for (size_t t = 0; t < num_t; t++) merge_thread(t);
#endif

  // Check merge errors
  size_t first_merge_error_line = 0;
  std::string first_merge_error_message;
  for (size_t t = 0; t < merge_error_lines.size(); t++) {
    if (merge_error_lines[t] == 0) continue;
    if (first_merge_error_line == 0 ||
        merge_error_lines[t] < first_merge_error_line) {
      first_merge_error_line = merge_error_lines[t];
      first_merge_error_message = merge_error_messages[t];
    }
  }

  if (first_merge_error_line != 0) {
    if (err) {
      (*err) += "Failed parse line(line " +
                std::to_string(first_merge_error_line) + "). " +
                first_merge_error_message + "\n";
    }
    return false;
  }

  // Handle vertex colors: only keep if all vertices have color
  {
    bool saw_any = false, saw_missing = false;
    for (size_t t = 0; t < num_t; t++) {
      saw_any = saw_any || (thread_saw_any_vertex_color[t] != 0);
      saw_missing = saw_missing || (thread_missing_vertex_color[t] != 0);
    }
    if (saw_any && !saw_missing) {
      attrib.colors = all_colors;  // share arena pointer
    }
    // else attrib.colors stays empty (default)
  }

  // Compute actual sizes and compact/truncate
  size_t actual_num_indices = 0;
  size_t actual_num_faces = 0;
  int greatest_v_idx = -1, greatest_vn_idx = -1, greatest_vt_idx = -1;
  for (size_t t = 0; t < num_t; t++) {
    actual_num_indices += written_index_counts[t];
    actual_num_faces += written_face_counts[t];
    opt_updateGreatestIndex(thread_greatest_v_idx[t], &greatest_v_idx);
    opt_updateGreatestIndex(thread_greatest_vn_idx[t], &greatest_vn_idx);
    opt_updateGreatestIndex(thread_greatest_vt_idx[t], &greatest_vt_idx);
  }

  // Compact face arrays to remove gaps left by threads that wrote fewer
  // indices/faces than pre-allocated (e.g. triangulation returning 0 for
  // faces with out-of-bounds vertex indices).
  if (num_t > 1) {
    size_t dst_idx = 0;
    size_t dst_face = 0;
    for (size_t t = 0; t < num_t; t++) {
      size_t src_idx = f_off[t];
      size_t src_face = face_off[t];
      size_t idx_count = written_index_counts[t];
      size_t fc_count = written_face_counts[t];
      if (dst_idx != src_idx && idx_count > 0) {
        std::memmove(&attrib.indices[dst_idx], &attrib.indices[src_idx],
                     idx_count * sizeof(index_t));
      }
      if (dst_face != src_face && fc_count > 0) {
        std::memmove(&attrib.face_num_verts[dst_face],
                     &attrib.face_num_verts[src_face],
                     fc_count * sizeof(int));
        std::memmove(&attrib.material_ids[dst_face],
                     &attrib.material_ids[src_face],
                     fc_count * sizeof(int));
        if (any_smoothing) {
          std::memmove(&all_smoothing_group_ids[dst_face],
                       &all_smoothing_group_ids[src_face],
                       fc_count * sizeof(unsigned int));
        }
      }
      dst_idx += idx_count;
      dst_face += fc_count;
    }
  }

  attrib.indices.truncate(actual_num_indices);
  attrib.face_num_verts.truncate(actual_num_faces);
  attrib.material_ids.truncate(actual_num_faces);
  if (any_smoothing) {
    all_smoothing_group_ids.truncate(actual_num_faces);
    attrib.smoothing_group_ids = all_smoothing_group_ids;
  }

  opt_appendOutOfBoundsWarnings(
      warn, greatest_v_idx, greatest_vn_idx, greatest_vt_idx,
      static_cast<int>(attrib.vertices.size() / 3),
      static_cast<int>(attrib.normals.size() / 3),
      static_cast<int>(attrib.texcoords.size() / 2),
      all_line_infos.size() + 1);
  for (size_t deg = 0; deg < eof_pending_degenerate_faces; deg++) {
    if (warn) (*warn) += "Degenerated face found\n.";
  }

  // ---- Phase 5: construct shape ranges (views, no copies) ----
  {
    const size_t total_faces = attrib.face_num_verts.size();
    // Build prefix-sum for index offsets
    TypedArray<size_t> idx_prefix;
    idx_prefix.allocate(arena, total_faces + 1);
    idx_prefix[0] = 0;
    for (size_t fi = 0; fi < total_faces; fi++) {
      idx_prefix[fi + 1] =
          idx_prefix[fi] + static_cast<size_t>(attrib.face_num_verts[fi]);
    }

    size_t face_count = 0;
    std::string current_name;
    size_t face_prev_offset = 0;
    bool have_active_shape_name = false;

    auto emit_shape = [&](const std::string &name, size_t f_start, size_t f_end) {
      if (f_end <= f_start) return;
      OptShapeRange sr;
      sr.name = name;
      sr.face_offset = f_start;
      sr.face_count = f_end - f_start;
      sr.index_offset = idx_prefix[f_start];
      sr.index_count = idx_prefix[f_end] - idx_prefix[f_start];
      result->shapes.push_back(std::move(sr));
    };

    for (size_t t = 0; t < num_t; t++) {
      const OptThreadData &td = thread_data[t];
      for (size_t si = 0; si < td.seq.size(); si++) {
        const OptSeqEntry &entry = td.seq[si];
        if (entry.kind == OptSeqEntry::SEQ_META) {
          const OptMetaCmd &meta = td.metas[entry.index];
          if (meta.type == OPT_CMD_O || meta.type == OPT_CMD_G) {
            std::string name;
            if (meta.type == OPT_CMD_O && meta.str_ptr) {
              name.assign(meta.str_ptr, meta.str_len);
            } else if (!meta.str_storage.empty()) {
              name = meta.str_storage;
            } else if (meta.str_ptr) {
              name.assign(meta.str_ptr, meta.str_len);
            }
            while (!name.empty() &&
                   (name.back() == '\r' || name.back() == '\n'))
              name.pop_back();

            if (face_count == 0) {
              current_name = name;
              face_prev_offset = 0;
              have_active_shape_name = true;
            } else {
              if (!have_active_shape_name) {
                emit_shape(std::string(), 0, face_count);
              } else if (face_count > face_prev_offset) {
                emit_shape(current_name, face_prev_offset, face_count);
              }
              current_name = name;
              face_prev_offset = face_count;
              have_active_shape_name = true;
            }
          }
        } else {
          // SEQ_FACE
          face_count += command_written_faces[t][entry.index];
        }
      }
    }

    // Final shape
    if (face_count > face_prev_offset || face_count == 0) {
      emit_shape(current_name, face_prev_offset, face_count);
    }
  }

  result->valid = true;
  return true;
}

// ---- LoadObjOptTyped (buffer version, public API) ----

OptResult LoadObjOptTyped(const char *buf, size_t buf_len,
                          std::string *warn, std::string *err,
                          const OptLoadConfig &config) {
  OptResult result;
  LoadObjOptTyped_internal(&result, warn, err, buf, buf_len,
                           std::string(), "<stream>", false, config);
  return result;
}

// ---- LoadObjOptTyped (file version) ----

OptResult LoadObjOptTyped(const char *filename,
                          std::string *warn, std::string *err,
                          const char *mtl_basedir,
                          const OptLoadConfig &config) {
  OptResult result;
  if (!filename) {
    if (err) *err = "filename is null.";
    return result;
  }

  std::string filepath(filename);
  std::string baseDir;
  if (mtl_basedir) {
    baseDir = mtl_basedir;
  } else {
    size_t pos = filepath.find_last_of("/\\");
    if (pos != std::string::npos) {
      baseDir = filepath.substr(0, pos + 1);
    }
  }
  if (!baseDir.empty()) {
#ifndef _WIN32
    const char dirsep = '/';
#else
    const char dirsep = '\\';
#endif
    if (baseDir[baseDir.length() - 1] != dirsep) baseDir += dirsep;
  }

#ifdef TINYOBJLOADER_USE_MMAP
  {
    MappedFile mf;
    if (mf.open(filepath.c_str())) {
      LoadObjOptTyped_internal(&result, warn, err, mf.data, mf.size,
                               baseDir, filepath, true, config);
      return result;
    }
  }
#endif

#ifdef _WIN32
  std::ifstream ifs(LongPathW(UTF8ToWchar(filepath)).c_str(),
                    std::ios::binary | std::ios::ate);
#else
  std::ifstream ifs(filepath.c_str(), std::ios::binary | std::ios::ate);
#endif
  if (!ifs.is_open()) {
    if (err) *err = "Cannot open file: " + filepath;
    return result;
  }

  std::streamsize fsize = ifs.tellg();
  ifs.seekg(0, std::ios::beg);

  if (fsize <= 0) {
    LoadObjOptTyped_internal(&result, warn, err, "", 0,
                             baseDir, filepath, true, config);
    return result;
  }

  std::vector<char> file_buf(static_cast<size_t>(fsize));
  if (!ifs.read(file_buf.data(), fsize)) {
    if (err) *err = "Failed to read file: " + filepath;
    return result;
  }
  ifs.close();

  LoadObjOptTyped_internal(&result, warn, err, file_buf.data(),
                           static_cast<size_t>(fsize), baseDir, filepath,
                           true, config);
  return result;
}

#ifdef __clang__
#pragma clang diagnostic pop
#endif
}  // namespace tinyobj

#endif