src/hb-subset-cff2.cc

/*
 * Copyright © 2018 Adobe Inc.
 *
 *  This is part of HarfBuzz, a text shaping library.
 *
 * Permission is hereby granted, without written agreement and without
 * license or royalty fees, to use, copy, modify, and distribute this
 * software and its documentation for any purpose, provided that the
 * above copyright notice and the following two paragraphs appear in
 * all copies of this software.
 *
 * IN NO EVENT SHALL THE COPYRIGHT HOLDER BE LIABLE TO ANY PARTY FOR
 * DIRECT, INDIRECT, SPECIAL, INCIDENTAL, OR CONSEQUENTIAL DAMAGES
 * ARISING OUT OF THE USE OF THIS SOFTWARE AND ITS DOCUMENTATION, EVEN
 * IF THE COPYRIGHT HOLDER HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH
 * DAMAGE.
 *
 * THE COPYRIGHT HOLDER SPECIFICALLY DISCLAIMS ANY WARRANTIES, INCLUDING,
 * BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
 * FITNESS FOR A PARTICULAR PURPOSE.  THE SOFTWARE PROVIDED HEREUNDER IS
 * ON AN "AS IS" BASIS, AND THE COPYRIGHT HOLDER HAS NO OBLIGATION TO
 * PROVIDE MAINTENANCE, SUPPORT, UPDATES, ENHANCEMENTS, OR MODIFICATIONS.
 *
 * Adobe Author(s): Michiharu Ariza
 */

#include "hb.hh"

#ifndef HB_NO_SUBSET_CFF

#include "hb-open-type.hh"
#include "hb-ot-cff2-table.hh"
#include "hb-ot-var-common.hh"
#include "hb-set.h"
#include "hb-subset-plan.hh"
#include "hb-subset-cff-common.hh"
#include "hb-cff2-interp-cs.hh"
#include "hb-subset-cff2-to-cff1.hh"

using namespace CFF;

namespace CFF {

/*
 * Partial instancing (some axes pinned or limited, others kept).
 *
 * This is a port of fontTools' instantiateCFF2 (fonttools/fonttools#3506),
 * with one simplification: the variation-store instancing solver acts on
 * regions independently of the per-blend delta values, so instead of lifting
 * every blend's deltas into the store, instancing it, and reading the rows
 * back, we instance each VarData once with identity delta rows.  The
 * instantiated tuples then give, per VarData, a fixed linear map from old
 * per-region deltas to new per-region deltas.  The contribution that folds
 * into each blend's default value is the old store's region scalars
 * evaluated at the new default location (plan->normalized_coords); the
 * surviving rebased regions all evaluate to zero there.
 *
 * Use: create (), then note_use () for every blend the retained charstrings
 * and private dicts hold, then finalize (); only VarDatas that are both
 * still variable and actually used survive into the new store.
 */
struct cff2_instancing_plan_t
{
  struct vardata_transform_t
  {
    /* new_deltas[j] = Σ_i matrix[j][i] * old_deltas[i] */
    hb_vector_t<hb_vector_t<float>> matrix;
    /* Default-value fold-in scalars, one per old region: the old store's
     * region scalars at the new default location. */
    hb_vector_t<float> gains;
    /* Surviving regions, parallel to matrix rows. */
    hb_vector_t<hb_hashmap_t<hb_tag_t, Triple>> regions;
    /* Index into the new global region list, per new region (finalize). */
    hb_vector_t<unsigned> region_indices;
    unsigned new_major = 0;
    bool alive_ = false;
    bool alive () const { return alive_; }

    /* One value's delta for new region j, from its old per-region deltas. */
    double transform_delta (unsigned j, hb_array_t<const number_t> deltas) const
    {
      const hb_vector_t<float> &col = matrix.arrayZ[j];
      double d = 0.;
      unsigned count = hb_min (col.length, deltas.length);
      for (unsigned c = 0; c < count; c++)
	d += (double) col.arrayZ[c] * deltas.arrayZ[c].to_real ();
      return d;
    }
  };

  hb_vector_t<vardata_transform_t> transforms;  /* per old VarData (vsindex) */
  hb_vector_t<hb_hashmap_t<hb_tag_t, Triple>> new_regions;  /* global region list */
  hb_set_t used_majors;
  unsigned new_major_count = 0;

  bool has_variations () const { return new_major_count; }

  const vardata_transform_t *get_transform (unsigned major) const
  { return major < transforms.length ? &transforms.arrayZ[major] : nullptr; }

  /* Record that a retained blend references this VarData. */
  void note_use (unsigned major) { used_majors.add (major); }

  /* The new vsindex a private dict's (or charstring's) old ivs maps to;
   * dead VarDatas fall back to 0, their blends having become static. */
  unsigned remap_ivs (unsigned old_ivs) const
  {
    const vardata_transform_t *t = get_transform (old_ivs);
    return (t && t->alive ()) ? t->new_major : 0;
  }

  double fold_deltas (unsigned old_ivs, hb_array_t<const number_t> deltas) const
  {
    const vardata_transform_t *t = get_transform (old_ivs);
    if (unlikely (!t || t->gains.length != deltas.length)) return 0.;
    double v = 0.;
    for (unsigned i = 0; i < deltas.length; i++)
      v += (double) t->gains.arrayZ[i] * deltas.arrayZ[i].to_real ();
    return v;
  }

  bool create (const CFF2ItemVariationStore *cff2VarStore,
	       const hb_subset_plan_t *plan)
  {
    const OT::ItemVariationStore &varStore = cff2VarStore->varStore;
    unsigned major_count = varStore.get_sub_table_count ();
    if (!transforms.resize (major_count)) return false;

    if (!varStore.get_region_list ().get_var_regions (plan->axes_old_index_tag_map, orig_regions))
      return false;

    /* Instance each VarData with identity delta rows; the surviving tuples'
     * delta columns are the old-region → new-region transform. */
    OT::optimize_scratch_t scratch;
    for (unsigned m = 0; m < major_count; m++)
    {
      const OT::VarData &var_data = varStore.get_sub_table (m);
      unsigned k = var_data.get_region_index_count ();
      vardata_transform_t &t = transforms.arrayZ[m];

      if (!t.gains.resize_exact (k)) return false;
      varStore.get_region_scalars (m, plan->normalized_coords.arrayZ,
				   plan->normalized_coords.length,
				   t.gains.arrayZ, k);

      OT::tuple_variations_t tuples;
      for (unsigned r = 0; r < k; r++)
      {
	OT::tuple_delta_t tuple;
	if (!tuple.deltas_x.resize (k) ||
	    !tuple.indices.resize (k))
	  return false;
	for (unsigned i = 0; i < k; i++)
	  tuple.indices.arrayZ[i] = true;
	tuple.deltas_x.arrayZ[r] = 1.f;
	unsigned region_index = var_data.get_region_index (r);
	if (unlikely (region_index >= orig_regions.length)) return false;
	tuple.axis_tuples = orig_regions.arrayZ[region_index];
	if (unlikely (tuple.axis_tuples.in_error ())) return false;
	tuples.tuple_vars.push (std::move (tuple));
      }
      if (unlikely (tuples.tuple_vars.in_error ())) return false;

      if (!tuples.instantiate (plan->axes_location, plan->axes_triple_distances, scratch))
	return false;

      for (auto &tuple : tuples.tuple_vars)
      {
	if (unlikely (tuple.deltas_x.length != k)) return false;
	hb_vector_t<float> col;
	if (unlikely (!col.resize_exact (k))) return false;
	for (unsigned i = 0; i < k; i++)
	  col.arrayZ[i] = tuple.deltas_x.arrayZ[i];
	t.matrix.push (std::move (col));
	t.regions.push (std::move (tuple.axis_tuples));
      }
      if (unlikely (t.matrix.in_error () || t.regions.in_error ())) return false;
    }

    return true;
  }

  /* Decide which VarDatas survive, number them, and build the new global
   * region list: surviving original regions first, in their original order,
   * then solver-split new regions in order of appearance (fontTools'
   * rebuildRegions + prune_regions). */
  bool finalize ()
  {
    for (unsigned m = 0; m < transforms.length; m++)
    {
      vardata_transform_t &t = transforms.arrayZ[m];
      t.alive_ = t.matrix.length && used_majors.has (m);
      if (t.alive_)
	t.new_major = new_major_count++;
    }

    /* Keys hash by content, so pointers into different owners compare fine. */
    hb_hashmap_t<const hb_hashmap_t<hb_tag_t, Triple>*, unsigned> region_idx_map;
    hb_vector_t<const hb_hashmap_t<hb_tag_t, Triple>*> region_ptrs;
    for (const vardata_transform_t &t : transforms)
    {
      if (!t.alive ()) continue;
      for (const auto &region : t.regions)
	if (!region_idx_map.has (&region))
	  if (unlikely (!region_idx_map.set (&region, (unsigned) -1)))
	    return false;
    }
    for (const auto &r : orig_regions)
    {
      unsigned *idx;
      if (region_idx_map.has (&r, &idx) && *idx == (unsigned) -1)
      {
	if (unlikely (!region_idx_map.set (&r, region_ptrs.length))) return false;
	region_ptrs.push (&r);
      }
    }
    for (const vardata_transform_t &t : transforms)
    {
      if (!t.alive ()) continue;
      for (const auto &region : t.regions)
      {
	unsigned *idx;
	if (region_idx_map.has (&region, &idx) && *idx == (unsigned) -1)
	{
	  if (unlikely (!region_idx_map.set (&region, region_ptrs.length))) return false;
	  region_ptrs.push (&region);
	}
      }
    }
    if (unlikely (region_ptrs.in_error ())) return false;

    for (vardata_transform_t &t : transforms)
    {
      if (!t.alive ()) continue;
      for (const auto &region : t.regions)
      {
	unsigned *idx;
	if (unlikely (!region_idx_map.has (&region, &idx))) return false;
	t.region_indices.push (*idx);
      }
      if (unlikely (t.region_indices.in_error ())) return false;
    }

    if (!new_regions.resize (region_ptrs.length)) return false;
    for (unsigned i = 0; i < region_ptrs.length; i++)
    {
      new_regions.arrayZ[i] = *region_ptrs.arrayZ[i];
      if (unlikely (new_regions.arrayZ[i].in_error ())) return false;
    }

    return true;
  }

  /* Serialize the instanced store: a CFF2ItemVariationStore whose VarDatas
   * carry no items, just region indexes (blend deltas live in charstrings).
   * Laid out manually since all sizes are known up front. */
  bool serialize_var_store (hb_serialize_context_t *c,
			    const hb_vector_t<hb_tag_t> &axis_tags) const
  {
    unsigned data_count = new_major_count;
    unsigned header_size = 2 + 4 + 2 + 4 * data_count;

    uint64_t offset = header_size;
    hb_vector_t<unsigned> data_offsets;
    for (const vardata_transform_t &t : transforms)
    {
      if (!t.alive ()) continue;
      data_offsets.push ((unsigned) offset);
      offset += 6 + 2 * (uint64_t) t.region_indices.length;
    }
    if (unlikely (data_offsets.in_error () || data_offsets.length != data_count))
      return false;
    unsigned regions_offset = (unsigned) offset;
    uint64_t total = offset + 4 + (uint64_t) new_regions.length * axis_tags.length * 6;
    if (unlikely (total > 0xFFFFu)) return false;

    HBUINT16 *size_field = c->allocate_size<HBUINT16> (HBUINT16::static_size);
    if (unlikely (!size_field)) return false;
    *size_field = total;

    char *header = c->allocate_size<char> (header_size);
    if (unlikely (!header)) return false;
    * (HBUINT16 *) (header + 0) = 1;			/* format */
    * (HBUINT32 *) (header + 2) = regions_offset;
    * (HBUINT16 *) (header + 6) = data_count;
    for (unsigned i = 0; i < data_count; i++)
      * (HBUINT32 *) (header + 8 + 4 * i) = data_offsets.arrayZ[i];

    for (const vardata_transform_t &t : transforms)
    {
      if (!t.alive ()) continue;
      unsigned k = t.region_indices.length;
      HBUINT16 *v = (HBUINT16 *) c->allocate_size<char> (6 + 2 * k);
      if (unlikely (!v)) return false;
      v[0] = 0;						/* itemCount */
      v[1] = 0;						/* wordSizeCount */
      v[2] = k;						/* regionIndices.len */
      for (unsigned j = 0; j < k; j++)
	v[3 + j] = t.region_indices.arrayZ[j];
    }

    hb_vector_t<const hb_hashmap_t<hb_tag_t, Triple>*> region_ptrs;
    if (unlikely (!region_ptrs.alloc_exact (new_regions.length))) return false;
    for (const auto &r : new_regions)
      region_ptrs.push (&r);
    auto *region_list = c->start_embed<OT::VarRegionList> ();
    return region_list->serialize (c, axis_tags, region_ptrs);
  }

  private:
  hb_vector_t<hb_hashmap_t<hb_tag_t, Triple>> orig_regions;
};

/* Usage-collection pass for partial instancing: record which VarDatas the
 * retained charstrings' and private dicts' blends reference.  Mirrors the
 * emission pass: blends consumed by dropped hint ops don't count. */
struct cff2_usage_param_t
{
  void init () {}

  cff2_instancing_plan_t *instancing = nullptr;
  bool drop_hints = false;
  unsigned ivs = 0;	/* dict-side current vsindex */
};

struct cff2_cs_opset_usage_t : cff2_cs_opset_t<cff2_cs_opset_usage_t, cff2_usage_param_t, blend_arg_t>
{
  static void flush_args_and_op (op_code_t op, cff2_cs_interp_env_t<blend_arg_t> &env, cff2_usage_param_t& param)
  {
    bool counts = true;
    switch (op)
    {
      case OpCode_return:
      case OpCode_endchar:
	counts = false;
	break;

      case OpCode_hstem:
      case OpCode_hstemhm:
      case OpCode_vstem:
      case OpCode_vstemhm:
      case OpCode_hintmask:
      case OpCode_cntrmask:
	if (param.drop_hints)
	  counts = false;
	break;

      default:
	break;
    }

    if (counts)
      for (unsigned int i = 0; i < env.argStack.get_count (); i++)
	if (env.argStack[i].blending ())
	{
	  param.instancing->note_use (env.get_ivs ());
	  break;
	}

    env.clear_args ();
  }
};

struct cff2_private_dict_usage_opset_t : dict_opset_t
{
  static void process_op (op_code_t op, cff2_priv_dict_interp_env_t& env, cff2_usage_param_t& param)
  {
    switch (op) {
      case OpCode_vsindexdict:
	env.process_vsindex ();
	param.ivs = env.get_ivs ();
	env.clear_args ();
	return;
      case OpCode_blenddict:
	param.instancing->note_use (param.ivs);
	env.clear_args ();
	return;
      default:
	dict_opset_t::process_op (op, env);
	if (!env.argStack.is_empty ()) return;
	env.clear_args ();
	return;
    }
  }
};

} /* namespace CFF */

struct cff2_sub_table_info_t : cff_sub_table_info_t
{
  cff2_sub_table_info_t ()
    : cff_sub_table_info_t (),
      var_store_link (0)
  {}

  objidx_t  var_store_link;
};

struct cff2_top_dict_op_serializer_t : cff_top_dict_op_serializer_t<>
{
  bool serialize (hb_serialize_context_t *c,
		  const op_str_t &opstr,
		  const cff2_sub_table_info_t &info) const
  {
    TRACE_SERIALIZE (this);

    switch (opstr.op)
    {
      case OpCode_vstore:
        if (info.var_store_link)
	  return_trace (FontDict::serialize_link4_op(c, opstr.op, info.var_store_link));
	else
	  return_trace (true);

      default:
	return_trace (cff_top_dict_op_serializer_t<>::serialize (c, opstr, info));
    }
  }
};

struct cff2_cs_opset_flatten_t : cff2_cs_opset_t<cff2_cs_opset_flatten_t, flatten_param_t, blend_arg_t>
{
  static void flush_args_and_op (op_code_t op, cff2_cs_interp_env_t<blend_arg_t> &env, flatten_param_t& param)
  {
    /* Optionally capture command for specialization (before flushing, to preserve args) */
    if (param.commands)
    {
      bool skip_command = false;

      switch (op)
      {
	case OpCode_return:
	case OpCode_endchar:
	  skip_command = true;
	  break;

	case OpCode_hstem:
	case OpCode_hstemhm:
	case OpCode_vstem:
	case OpCode_vstemhm:
	case OpCode_hintmask:
	case OpCode_cntrmask:
	  if (param.drop_hints)
	    skip_command = true;
	  break;

	default:
	  break;
      }

      if (!skip_command)
      {
	cs_command_t cmd (op);
	/* Capture resolved blend values */
	for (unsigned int i = 0; i < env.argStack.get_count ();)
	{
	  const blend_arg_t &arg = env.argStack[i];
	  if (arg.blending ())
	  {
	    /* For blend args, capture only the resolved default value */
	    cmd.args.push (arg);
	    /* Skip over the multiple blend values */
	    i += arg.numValues;
	  }
	  else
	  {
	    cmd.args.push (arg);
	    i++;
	  }
	}
	param.commands->push (cmd);
      }
    }

    switch (op)
    {
      case OpCode_return:
      case OpCode_endchar:
	/* dummy opcodes in CFF2. ignore */
	if (op == OpCode_endchar && param.instancer)
	  emit_vsindex_prefix (env, param);
	break;

      case OpCode_hstem:
      case OpCode_hstemhm:
      case OpCode_vstem:
      case OpCode_vstemhm:
      case OpCode_hintmask:
      case OpCode_cntrmask:
	if (param.drop_hints)
	{
	  env.clear_args ();
	  return;
	}
	HB_FALLTHROUGH;

      default:
	SUPER::flush_args_and_op (op, env, param);
	break;
    }
  }

  static void flush_args (cff2_cs_interp_env_t<blend_arg_t> &env, flatten_param_t& param)
  {
    for (unsigned int i = 0; i < env.argStack.get_count ();)
    {
      const blend_arg_t &arg = env.argStack[i];
      if (arg.blending ())
      {
	if (unlikely (!((arg.numValues > 0) && (env.argStack.get_count () >= arg.numValues))))
	{
	  env.set_error ();
	  return;
	}
	if (param.instancer)
	  rewrite_blends (arg, i, env, param);
	else
	  flatten_blends (arg, i, env, param);
	i += arg.numValues;
      }
      else
      {
	str_encoder_t  encoder (param.flatStr);
	encoder.encode_num_cs (arg);
	i++;
      }
    }
    SUPER::flush_args (env, param);
  }

  static void flatten_blends (const blend_arg_t &arg, unsigned int i, cff2_cs_interp_env_t<blend_arg_t> &env, flatten_param_t& param)
  {
    /* flatten the default values */
    str_encoder_t  encoder (param.flatStr);
    for (unsigned int j = 0; j < arg.numValues; j++)
    {
      const blend_arg_t &arg1 = env.argStack[i + j];
      if (unlikely (!((arg1.blending () && (arg.numValues == arg1.numValues) && (arg1.valueIndex == j) &&
	      (arg1.deltas.length == env.get_region_count ())))))
      {
	env.set_error ();
	return;
      }
      encoder.encode_num_cs (arg1);
    }
    /* flatten deltas for each value */
    for (unsigned int j = 0; j < arg.numValues; j++)
    {
      const blend_arg_t &arg1 = env.argStack[i + j];
      for (unsigned int k = 0; k < arg1.deltas.length; k++)
	encoder.encode_num_cs (arg1.deltas[k]);
    }
    /* flatten the number of values followed by blend operator */
    encoder.encode_int (arg.numValues);
    encoder.encode_op (OpCode_blendcs);
  }

  /* Partial instancing: rewrite one blend group against the instanced store.
   * Each value's default absorbs the folded (pinned-away) contribution, and
   * its deltas are mapped onto the new region list. */
  static void rewrite_blends (const blend_arg_t &arg, unsigned int i, cff2_cs_interp_env_t<blend_arg_t> &env, flatten_param_t& param)
  {
    const cff2_instancing_plan_t &instancer = *param.instancer;
    unsigned ivs = env.get_ivs ();
    const cff2_instancing_plan_t::vardata_transform_t *t = instancer.get_transform (ivs);
    if (unlikely (!t))
    {
      env.set_error ();
      return;
    }

    unsigned n = arg.numValues;
    for (unsigned int j = 0; j < n; j++)
    {
      const blend_arg_t &arg1 = env.argStack[i + j];
      if (unlikely (!(arg1.blending () && (arg1.numValues == n) && (arg1.valueIndex == j) &&
		      (arg1.deltas.length == env.get_region_count ()) &&
		      (arg1.deltas.length == t->gains.length))))
      {
	env.set_error ();
	return;
      }
    }

    str_encoder_t encoder (param.flatStr);
    unsigned k_new = t->alive () ? t->matrix.length : 0;

    if (!k_new)
    {
      /* All of this VarData's regions died; blends dissolve into statics. */
      for (unsigned int j = 0; j < n; j++)
      {
	const blend_arg_t &arg1 = env.argStack[i + j];
	number_t v;
	v.set_real (arg1.to_real () + round (instancer.fold_deltas (ivs, arg1.deltas.as_array ())));
	encoder.encode_num_cs (v);
      }
      return;
    }

    /* Emit in chunks so the output charstring stays within the argument
     * stack limit: values already emitted for this op, plus this chunk's
     * defaults, deltas and count, must not exceed 513. */
    unsigned done = 0;
    while (done < n)
    {
      unsigned avail = 512 > (i + done) ? 512 - (i + done) : 0;
      unsigned chunk = hb_min (avail / (k_new + 1), n - done);
      if (unlikely (!chunk))
      {
	env.set_error ();
	return;
      }

      for (unsigned int j = done; j < done + chunk; j++)
      {
	const blend_arg_t &arg1 = env.argStack[i + j];
	number_t v;
	v.set_real (arg1.to_real () + round (instancer.fold_deltas (ivs, arg1.deltas.as_array ())));
	encoder.encode_num_cs (v);
      }
      for (unsigned int j = done; j < done + chunk; j++)
      {
	const blend_arg_t &arg1 = env.argStack[i + j];
	for (unsigned r = 0; r < k_new; r++)
	{
	  number_t v;
	  v.set_int (hb_clamp_to<int> (round (t->transform_delta (r, arg1.deltas.as_array ()))));
	  encoder.encode_num_cs (v);
	}
      }
      encoder.encode_int (chunk);
      encoder.encode_op (OpCode_blendcs);
      done += chunk;
    }
    param.emitted_blend = true;
  }

  /* Partial instancing: prepend a vsindex op when the charstring's
   * (remapped) ivs differs from its FD's (remapped) private-dict ivs. */
  static void emit_vsindex_prefix (cff2_cs_interp_env_t<blend_arg_t> &env, flatten_param_t& param)
  {
    if (!param.emitted_blend) return;
    const cff2_instancing_plan_t &instancer = *param.instancer;
    unsigned new_ivs = instancer.remap_ivs (env.get_ivs ());
    unsigned fd_ivs = instancer.remap_ivs (env.get_orig_ivs ());
    if (new_ivs == fd_ivs) return;

    str_buff_t prefix;
    str_encoder_t enc (prefix);
    enc.encode_int ((int) new_ivs);
    enc.encode_op (OpCode_vsindexcs);

    str_buff_t &buf = param.flatStr;
    unsigned old_len = buf.length;
    if (unlikely (prefix.in_error () || !buf.resize (old_len + prefix.length)))
    {
      env.set_error ();
      return;
    }
    memmove (buf.arrayZ + prefix.length, buf.arrayZ, old_len);
    hb_memcpy (buf.arrayZ, prefix.arrayZ, prefix.length);
  }

  static void flush_op (op_code_t op, cff2_cs_interp_env_t<blend_arg_t> &env, flatten_param_t& param)
  {
    switch (op)
    {
      case OpCode_return:
      case OpCode_endchar:
	return;
      default:
	str_encoder_t  encoder (param.flatStr);
	encoder.encode_op (op);
    }
  }

  static void flush_hintmask (op_code_t op, cff2_cs_interp_env_t<blend_arg_t> &env, flatten_param_t& param)
  {
    SUPER::flush_hintmask (op, env, param);
    /* Preserve hintmask payload in captured commands for specializer re-encoding. */
    if (param.commands && !param.drop_hints && param.commands->length > 0)
    {
      auto &cmd = param.commands->tail ();
      if (cmd.op == op)
      {
        cmd.mask_bytes.resize (env.hintmask_size);
        if (unlikely (cmd.mask_bytes.in_error ()))
        {
          env.set_error ();
          return;
        }
        for (unsigned int i = 0; i < env.hintmask_size; i++)
          cmd.mask_bytes[i] = env.str_ref[i];
      }
    }
    if (!param.drop_hints)
    {
      str_encoder_t  encoder (param.flatStr);
      for (unsigned int i = 0; i < env.hintmask_size; i++)
	encoder.encode_byte (env.str_ref[i]);
    }
  }

  private:
  typedef cff2_cs_opset_t<cff2_cs_opset_flatten_t, flatten_param_t, blend_arg_t> SUPER;
  typedef cs_opset_t<blend_arg_t, cff2_cs_opset_flatten_t, cff2_cs_opset_flatten_t, cff2_cs_interp_env_t<blend_arg_t>, flatten_param_t> CSOPSET;
};

struct cff2_cs_opset_subr_subset_t : cff2_cs_opset_t<cff2_cs_opset_subr_subset_t, subr_subset_param_t, blend_arg_t>
{
  static void process_op (op_code_t op, cff2_cs_interp_env_t<blend_arg_t> &env, subr_subset_param_t& param)
  {
    switch (op) {

      case OpCode_return:
	param.current_parsed_str->flush_coalesced (env.str_ref);
	param.current_parsed_str->set_parsed ();
	env.return_from_subr ();
	param.set_current_str (env, false);
	break;

      case OpCode_endchar:
	param.current_parsed_str->flush_coalesced (env.str_ref);
	param.current_parsed_str->set_parsed ();
	SUPER::process_op (op, env, param);
	break;

      case OpCode_callsubr:
	process_call_subr (op, CSType_LocalSubr, env, param, env.localSubrs, param.local_closure);
	break;

      case OpCode_callgsubr:
	process_call_subr (op, CSType_GlobalSubr, env, param, env.globalSubrs, param.global_closure);
	break;

      default:
	SUPER::process_op (op, env, param);
	param.current_parsed_str->add_op (op, env.str_ref);
	break;
    }
  }

  protected:
  static void process_call_subr (op_code_t op, cs_type_t type,
				 cff2_cs_interp_env_t<blend_arg_t> &env, subr_subset_param_t& param,
				 cff2_biased_subrs_t& subrs, hb_set_t *closure)
  {
    byte_str_ref_t    str_ref = env.str_ref;
    env.call_subr (subrs, type);
    param.current_parsed_str->add_call_op (op, str_ref, env.context.subr_num);
    closure->add (env.context.subr_num);
    param.set_current_str (env, true);
  }

  private:
  typedef cff2_cs_opset_t<cff2_cs_opset_subr_subset_t, subr_subset_param_t, blend_arg_t> SUPER;
};

struct cff2_subr_subsetter_t : subr_subsetter_t<cff2_subr_subsetter_t, CFF2Subrs, const OT::cff2::accelerator_subset_t, cff2_cs_interp_env_t<blend_arg_t>, cff2_cs_opset_subr_subset_t>
{
  cff2_subr_subsetter_t (const OT::cff2::accelerator_subset_t &acc_, const hb_subset_plan_t *plan_)
    : subr_subsetter_t (acc_, plan_) {}

  static void complete_parsed_str (cff2_cs_interp_env_t<blend_arg_t> &env, subr_subset_param_t& param, parsed_cs_str_t &charstring)
  {
    /* vsindex is inserted at the beginning of the charstring as necessary */
    if (env.seen_vsindex ())
    {
      number_t  ivs;
      ivs.set_int ((int)env.get_ivs ());
      charstring.set_prefix (ivs, OpCode_vsindexcs);
    }
  }
};

struct cff2_private_blend_encoder_param_t
{
  cff2_private_blend_encoder_param_t (hb_serialize_context_t *c,
				      const CFF2ItemVariationStore *varStore,
				      hb_array_t<int> normalized_coords) :
    c (c), varStore (varStore), normalized_coords (normalized_coords) {}

  void init () {}

  void process_blend ()
  {
    if (!seen_blend)
    {
      region_count = varStore->varStore.get_region_index_count (ivs);
      if (unlikely (!scalars.resize_exact (region_count)))
      {
	if (c) c->err (HB_SERIALIZE_ERROR_OTHER);
	seen_blend = true;
	return;
      }
      varStore->varStore.get_region_scalars (ivs, normalized_coords.arrayZ, normalized_coords.length,
					     &scalars[0], region_count);
      seen_blend = true;
    }
  }

  double blend_deltas (hb_array_t<const number_t> deltas) const
  {
    double v = 0;
    if (likely (scalars.length == deltas.length))
    {
      unsigned count = scalars.length;
      for (unsigned i = 0; i < count; i++)
	v += (double) scalars.arrayZ[i] * deltas.arrayZ[i].to_real ();
    }
    return v;
  }


  hb_serialize_context_t *c = nullptr;
  bool seen_blend = false;
  unsigned ivs = 0;
  unsigned region_count = 0;
  hb_vector_t<float> scalars;
  const	 CFF2ItemVariationStore *varStore = nullptr;
  hb_array_t<int> normalized_coords;

  /* CFF2 partial instancing: when set, blends are rewritten against the
   * instanced variation store instead of being flattened. */
  const cff2_instancing_plan_t *instancer = nullptr;
  /* Error-compensated fold rounding state, per dict operator. */
  double fold_exact = 0.;
  double fold_emitted = 0.;
};

struct cff2_private_dict_blend_opset_t : dict_opset_t
{
  static void process_arg_blend (cff2_private_blend_encoder_param_t& param,
				 number_t &arg,
				 const hb_array_t<const number_t> blends,
				 unsigned n, unsigned i)
  {
    arg.set_int (round (arg.to_real () + param.blend_deltas (blends)));
  }

  /* The fold added to a value, error-compensated across the operator: most
   * blended private-dict values (BlueValues, StemSnap*, ...) are
   * delta-encoded, so independently rounded folds would accumulate in the
   * decoded absolutes; keeping the emitted cumulative fold equal to the
   * rounded exact cumulative fold bounds every absolute within ±0.5. */
  static double fold (cff2_private_blend_encoder_param_t& param,
		      const hb_array_t<const number_t> blends)
  {
    double f = param.blend_deltas (blends);
    double emit = round (param.fold_exact + f) - param.fold_emitted;
    param.fold_exact += f;
    param.fold_emitted += emit;
    return emit;
  }

  /* Partial instancing: emit the args preceding this blend group, then the
   * group itself rewritten against the instanced store, and clear the
   * stack.  The op's remaining args are flushed at process_op time. */
  static void rewrite_blends (cff2_priv_dict_interp_env_t& env,
			      cff2_private_blend_encoder_param_t& param,
			      unsigned start, unsigned n, unsigned k)
  {
    const cff2_instancing_plan_t &instancer = *param.instancer;
    const cff2_instancing_plan_t::vardata_transform_t *t = instancer.get_transform (param.ivs);
    if (unlikely (!t || t->gains.length != k))
    {
      env.set_error ();
      return;
    }

    str_buff_t str;
    str_encoder_t encoder (str);

    for (unsigned idx = 0; idx < start; idx++)
      encoder.encode_num_tp (env.argStack[idx]);

    unsigned k_new = t->alive () ? t->matrix.length : 0;
    if (!k_new)
    {
      /* This VarData died (or is unused by any retained charstring, which
       * can happen for FDs kept only for retain-gids holes); blends
       * dissolve into statics. */
      for (unsigned j = 0; j < n; j++)
      {
	const hb_array_t<const number_t> blends = env.argStack.sub_array (start + n + (j * k), k);
	number_t v;
	v.set_real (env.argStack[start + j].to_real () + fold (param, blends));
	encoder.encode_num_tp (v);
      }
    }
    else
    {
      unsigned done = 0;
      while (done < n)
      {
	unsigned avail = 512 > (start + done) ? 512 - (start + done) : 0;
	unsigned chunk = hb_min (avail / (k_new + 1), n - done);
	if (unlikely (!chunk))
	{
	  env.set_error ();
	  return;
	}

	for (unsigned j = done; j < done + chunk; j++)
	{
	  const hb_array_t<const number_t> blends = env.argStack.sub_array (start + n + (j * k), k);
	  number_t v;
	  v.set_real (env.argStack[start + j].to_real () + fold (param, blends));
	  encoder.encode_num_tp (v);
	}
	for (unsigned j = done; j < done + chunk; j++)
	{
	  const hb_array_t<const number_t> blends = env.argStack.sub_array (start + n + (j * k), k);
	  for (unsigned r = 0; r < k_new; r++)
	  {
	    number_t v;
	    v.set_int (hb_clamp_to<int> (round (t->transform_delta (r, blends))));
	    encoder.encode_num_tp (v);
	  }
	}
	encoder.encode_int (chunk);
	encoder.encode_op (OpCode_blenddict);
	done += chunk;
      }
    }

    auto bytes = str.as_bytes ();
    param.c->embed (&bytes, bytes.length);

    env.clear_args ();
  }

  static void process_blend (cff2_priv_dict_interp_env_t& env, cff2_private_blend_encoder_param_t& param)
  {
    unsigned int n, k;

    param.process_blend ();
    k = param.region_count;
    n = env.argStack.pop_uint ();

    unsigned count = env.argStack.get_count ();
    unsigned int total;
    if (unlikely (hb_unsigned_mul_overflows (k + 1, n, &total) || total > count))
    {
      env.set_error ();
      return;
    }

    /* copy the blend values into blend array of the default values */
    unsigned int start = count - total;
    /* let an obvious error case fail, but note CFF2 spec doesn't forbid n==0 */
    if (unlikely (start > count))
    {
      env.set_error ();
      return;
    }

    if (param.instancer && k)
    {
      rewrite_blends (env, param, start, n, k);
      return;
    }

    for (unsigned int i = 0; i < n; i++)
    {
      const hb_array_t<const number_t> blends = env.argStack.sub_array (start + n + (i * k), k);
      process_arg_blend (param, env.argStack[start + i], blends, n, i);
    }

    /* pop off blend values leaving default values now adorned with blend values */
    env.argStack.pop (k * n);
  }

  static void process_op (op_code_t op, cff2_priv_dict_interp_env_t& env, cff2_private_blend_encoder_param_t& param)
  {
    switch (op) {
      case OpCode_StdHW:
      case OpCode_StdVW:
      case OpCode_BlueScale:
      case OpCode_BlueShift:
      case OpCode_BlueFuzz:
      case OpCode_ExpansionFactor:
      case OpCode_LanguageGroup:
      case OpCode_BlueValues:
      case OpCode_OtherBlues:
      case OpCode_FamilyBlues:
      case OpCode_FamilyOtherBlues:
      case OpCode_StemSnapH:
      case OpCode_StemSnapV:
	break;
      case OpCode_vsindexdict:
	env.process_vsindex ();
	param.ivs = env.get_ivs ();
	if (param.instancer)
	{
	  /* Re-emit with the remapped vsindex; 0 is the dict default. */
	  unsigned new_ivs = param.instancer->remap_ivs (param.ivs);
	  if (new_ivs)
	  {
	    str_buff_t str;
	    str_encoder_t encoder (str);
	    encoder.encode_int ((int) new_ivs);
	    encoder.encode_op (OpCode_vsindexdict);
	    auto bytes = str.as_bytes ();
	    param.c->embed (&bytes, bytes.length);
	  }
	}
	env.clear_args ();
	return;
      case OpCode_blenddict:
	process_blend (env, param);
	return;

      default:
	dict_opset_t::process_op (op, env);
	if (!env.argStack.is_empty ()) return;
	break;
    }

    if (unlikely (env.in_error ())) return;

    // Write args then op

    str_buff_t str;
    str_encoder_t encoder (str);

    unsigned count = env.argStack.get_count ();
    for (unsigned i = 0; i < count; i++)
      encoder.encode_num_tp (env.argStack[i]);

    encoder.encode_op (op);

    auto bytes = str.as_bytes ();
    param.c->embed (&bytes, bytes.length);

    env.clear_args ();
    param.fold_exact = param.fold_emitted = 0.;
  }
};

struct cff2_private_dict_op_serializer_t : op_serializer_t
{
  cff2_private_dict_op_serializer_t (bool desubroutinize_, bool drop_hints_, bool pinned_,
				     const CFF::CFF2ItemVariationStore* varStore_,
				     hb_array_t<int> normalized_coords_,
				     const cff2_instancing_plan_t *instancer_ = nullptr)
    : desubroutinize (desubroutinize_), drop_hints (drop_hints_), pinned (pinned_),
      param (nullptr, varStore_, normalized_coords_)
  { param.instancer = instancer_; }

  bool serialize (hb_serialize_context_t *c,
		  const op_str_t &opstr,
		  objidx_t subrs_link) const
  {
    TRACE_SERIALIZE (this);

    if (drop_hints && dict_opset_t::is_hint_op (opstr.op))
      return_trace (true);

    if (opstr.op == OpCode_Subrs)
    {
      if (desubroutinize || !subrs_link)
	return_trace (true);
      else
	return_trace (FontDict::serialize_link2_op (c, opstr.op, subrs_link));
    }

    if (pinned || param.instancer)
    {
      /* Reinterpret opstr and process blends.  The param persists across
       * this dict's opstrs, so that an ivs set by a vsindexdict opstr is
       * still in effect when a later opstr's blends are processed. */
      cff2_priv_dict_interp_env_t env {hb_ubytes_t (opstr.ptr, opstr.length)};
      param.c = c;
      dict_interpreter_t<cff2_private_dict_blend_opset_t, cff2_private_blend_encoder_param_t, cff2_priv_dict_interp_env_t> interp (env);
      return_trace (interp.interpret (param));
    }

    return_trace (copy_opstr (c, opstr));
  }

  protected:
  const bool desubroutinize;
  const bool drop_hints;
  const bool pinned;
  mutable cff2_private_blend_encoder_param_t param;
};



namespace OT {
struct cff2_subset_plan
{
  bool create (const OT::cff2::accelerator_subset_t &acc,
	      hb_subset_plan_t *plan)
  {
    /* make sure notdef is first */
    hb_codepoint_t old_glyph;
    if (!plan->old_gid_for_new_gid (0, &old_glyph) || (old_glyph != 0)) return false;

    num_glyphs = plan->num_output_glyphs ();
    orig_fdcount = acc.fdArray->count;

    drop_hints = plan->flags & HB_SUBSET_FLAGS_NO_HINTING;
    bool instancing_requested = (bool) plan->normalized_coords;
    pinned = instancing_requested && plan->all_axes_pinned;
    normalized_coords = plan->normalized_coords;
    head_maxp_info = plan->head_maxp_info;
    hmtx_map = &plan->hmtx_map;
    desubroutinize = plan->flags & HB_SUBSET_FLAGS_DESUBROUTINIZE ||
		     instancing_requested; // For instancing we need this path

    /* Partial instancing: some axes pinned or limited, others kept. */
    partial_instancing = instancing_requested && !pinned &&
			 acc.varStore != &Null (CFF2ItemVariationStore);
    if (partial_instancing &&
	unlikely (!instancing.create (acc.varStore, plan)))
      return false;
    axis_tags = &plan->axis_tags;

    /* Enable command capture for CFF2→CFF1 conversion (for specialization) */
    capture_commands = pinned;

 #ifdef HB_EXPERIMENTAL_API
    min_charstrings_off_size = (plan->flags & HB_SUBSET_FLAGS_IFTB_REQUIREMENTS) ? 4 : 0;
 #else
    min_charstrings_off_size = 0;
 #endif

    if (desubroutinize)
    {
      if (partial_instancing)
      {
	/* Pass 1: collect which VarDatas the retained blends reference, so
	 * unused ones get pruned and the survivors renumbered. */
	hb_set_t used_fds;
	for (unsigned int i = 0; i < num_glyphs; i++)
	{
	  hb_codepoint_t glyph;
	  if (!plan->old_gid_for_new_gid (i, &glyph))
	    continue;
	  const hb_ubytes_t str = (*acc.charStrings)[glyph];
	  unsigned int fd = CFF::subset_get_fd (acc, glyph);
	  if (unlikely (fd >= acc.fdCount))
	    return false;
	  used_fds.add (fd);

	  cff2_cs_interp_env_t<blend_arg_t> env (str, acc, fd);
	  cs_interpreter_t<cff2_cs_interp_env_t<blend_arg_t>, cff2_cs_opset_usage_t, cff2_usage_param_t> interp (env);
	  cff2_usage_param_t param;
	  param.instancing = &instancing;
	  param.drop_hints = drop_hints;
	  if (unlikely (!interp.interpret (param)))
	    return false;
	}
	for (unsigned fd : used_fds)
	{
	  cff2_usage_param_t param;
	  param.instancing = &instancing;
	  unsigned count = acc.privateDicts[fd].get_count ();
	  for (unsigned j = 0; j < count; j++)
	  {
	    const op_str_t &opstr = acc.privateDicts[fd][j];
	    if (drop_hints && dict_opset_t::is_hint_op (opstr.op))
	      continue;
	    cff2_priv_dict_interp_env_t env {hb_ubytes_t (opstr.ptr, opstr.length)};
	    dict_interpreter_t<cff2_private_dict_usage_opset_t, cff2_usage_param_t, cff2_priv_dict_interp_env_t> interp (env);
	    if (unlikely (!interp.interpret (param)))
	      return false;
	  }
	}
	if (unlikely (!instancing.finalize ()))
	  return false;

	/* Pass 2: like subr_flattener_t, but blends stay symbolic (no coords
	 * fed to the env) and get rewritten against the instanced store. */
	if (!subset_charstrings.resize_exact (num_glyphs))
	  return false;
	for (unsigned int i = 0; i < num_glyphs; i++)
	{
	  hb_codepoint_t glyph;
	  if (!plan->old_gid_for_new_gid (i, &glyph))
	    continue;
	  const hb_ubytes_t str = (*acc.charStrings)[glyph];
	  unsigned int fd = CFF::subset_get_fd (acc, glyph);
	  if (unlikely (fd >= acc.fdCount))
	    return false;

	  cff2_cs_interp_env_t<blend_arg_t> env (str, acc, fd);
	  cs_interpreter_t<cff2_cs_interp_env_t<blend_arg_t>, cff2_cs_opset_flatten_t, flatten_param_t> interp (env);
	  flatten_param_t param = {
	    subset_charstrings.arrayZ[i],
	    (bool) (plan->flags & HB_SUBSET_FLAGS_NO_HINTING),
	    plan
	  };
	  param.instancer = &instancing;
	  if (unlikely (!interp.interpret (param)))
	    return false;
	}
      }
      else
      {
      /* Flatten global & local subrs */
      subr_flattener_t<const OT::cff2::accelerator_subset_t, cff2_cs_interp_env_t<blend_arg_t>, cff2_cs_opset_flatten_t>
		    flattener(acc, plan);

      /* Enable command capture if requested (for specialization) */
      if (capture_commands)
      {
	if (!charstring_commands.resize_exact (num_glyphs))
	  return false;

	if (!flattener.flatten (subset_charstrings, &charstring_commands))
	  return false;
      }
      else
      {
	if (!flattener.flatten (subset_charstrings))
	  return false;
      }
      }
    }
    else
    {
      cff2_subr_subsetter_t	subr_subsetter (acc, plan);

      /* Subset subrs: collect used subroutines, leaving all unused ones behind */
      if (!subr_subsetter.subset ())
	return false;

      /* encode charstrings, global subrs, local subrs with new subroutine numbers */
      if (!subr_subsetter.encode_charstrings (subset_charstrings, !pinned))
	return false;

      if (!subr_subsetter.encode_globalsubrs (subset_globalsubrs))
	return false;

      /* local subrs */
      if (!subset_localsubrs.resize (orig_fdcount))
	return false;
      for (unsigned int fd = 0; fd < orig_fdcount; fd++)
      {
	subset_localsubrs[fd].init ();
	if (!subr_subsetter.encode_localsubrs (fd, subset_localsubrs[fd]))
	  return false;
      }
    }

    /* FDSelect */
    if (acc.fdSelect != &Null (CFF2FDSelect))
    {
      if (unlikely (!hb_plan_subset_cff_fdselect (plan,
						  orig_fdcount,
						  *(const FDSelect *)acc.fdSelect,
						  acc.num_glyphs,
						  subset_fdcount,
						  subset_fdselect_size,
						  subset_fdselect_format,
						  subset_fdselect_ranges,
						  fdmap)))
	return false;
    }
    else
      fdmap.identity (1);

    return true;
  }

  cff2_sub_table_info_t info;

  unsigned int    num_glyphs;
  unsigned int    orig_fdcount = 0;
  unsigned int    subset_fdcount = 1;
  unsigned int    subset_fdselect_size = 0;
  unsigned int    subset_fdselect_format = 0;
  bool            pinned = false;
  hb_vector_t<code_pair_t>   subset_fdselect_ranges;

  hb_inc_bimap_t   fdmap;

  str_buff_vec_t	    subset_charstrings;
  str_buff_vec_t	    subset_globalsubrs;
  hb_vector_t<str_buff_vec_t> subset_localsubrs;

  bool	    drop_hints = false;
  bool	    desubroutinize = false;

  bool	    partial_instancing = false;
  cff2_instancing_plan_t instancing;
  const hb_vector_t<hb_tag_t> *axis_tags = nullptr;

  unsigned  min_charstrings_off_size = 0;

  hb_array_t<int> normalized_coords; // For instantiation
  head_maxp_info_t head_maxp_info;  // For FontBBox
  const hb_hashmap_t<hb_codepoint_t, hb_pair_t<unsigned, int>> *hmtx_map; // For widths

  // Width optimization results (for CFF1 conversion)
  unsigned default_width = 0;
  unsigned nominal_width = 0;

  // Command capture for specialization (CFF2→CFF1 conversion)
  bool capture_commands = false;
  hb_vector_t<hb_vector_t<cs_command_t>> charstring_commands;
};
} // namespace OT

/*
 * CFF2 to CFF1 Converter Implementation
 */

#include "hb-cff-width-optimizer.hh"
#include "hb-cff-specializer.hh"

/* Serialize charstrings using CFF1 format with widths */
static bool
_serialize_cff1_charstrings (hb_serialize_context_t *c,
                             OT::cff2_subset_plan &plan,
                             unsigned default_width,
                             unsigned nominal_width)
{
  c->push ();

  // CFF1 requires:
  // 1. Width at the beginning (if != defaultWidthX)
  // 2. endchar at the end
  str_buff_vec_t cff1_charstrings;
  if (unlikely (!cff1_charstrings.resize (plan.subset_charstrings.length)))
  {
    c->pop_discard ();
    return false;
  }

  for (unsigned i = 0; i < plan.subset_charstrings.length; i++)
  {
    // Get width for this glyph from hmtx_map
    unsigned width = 0;
    if (plan.hmtx_map->has (i))
      width = plan.hmtx_map->get (i).first;

    // Encode width if different from default
    str_encoder_t encoder (cff1_charstrings[i]);
    if (width != default_width)
    {
      int delta = (int) width - (int) nominal_width;
      encoder.encode_int (delta);
    }

    // Use specialized commands if available, otherwise use binary
    if (plan.capture_commands && i < plan.charstring_commands.length &&
        plan.charstring_commands[i].length > 0)
    {
      // Specialize and encode commands
      auto &commands = plan.charstring_commands[i];
      CFF::specialize_commands (commands, 48);  /* maxstack=48 for CFF1 */
      if (unlikely (!CFF::encode_commands (commands, cff1_charstrings[i])))
      {
        c->pop_discard ();
        return false;
      }
    }
    else
    {
      // Use binary CharString
      const str_buff_t &cs = plan.subset_charstrings[i];
      for (unsigned j = 0; j < cs.length; j++)
        cff1_charstrings[i].push (cs[j]);
    }

    // Check if it already ends with endchar (0x0e) or return (0x0b)
    if (cff1_charstrings[i].length == 0 ||
        (cff1_charstrings[i].tail () != 0x0e && cff1_charstrings[i].tail () != 0x0b))
    {
      // Append endchar operator
      if (unlikely (!cff1_charstrings[i].push_or_fail (0x0e)))
      {
        c->pop_discard ();
        return false;
      }
    }
  }

  unsigned data_size = 0;
  unsigned total_size = CFF1CharStrings::total_size (cff1_charstrings, &data_size, plan.min_charstrings_off_size);
  if (unlikely (!c->start_zerocopy (total_size)))
  {
    c->pop_discard ();
    return false;
  }

  auto *cs = c->start_embed<CFF1CharStrings> ();
  if (unlikely (!cs->serialize (c, cff1_charstrings, &data_size, plan.min_charstrings_off_size)))
  {
    c->pop_discard ();
    return false;
  }

  plan.info.char_strings_link = c->pop_pack (false);
  return true;
}

/* Serialize CID Charset (format 2 range: gid 0-N -> cid 0-N) */
static bool
_serialize_cff1_charset (hb_serialize_context_t *c,
                         unsigned int num_glyphs,
                         objidx_t &charset_link)
{
  // For CID fonts, create a simple identity charset
  // Format 2: one range covering all glyphs (except .notdef)
  c->push ();

  auto *charset = c->start_embed<Charset> ();
  if (unlikely (!charset))
  {
    c->pop_discard ();
    return false;
  }

  // Create a single range for CID 1 to num_glyphs-1
  hb_vector_t<code_pair_t> ranges;
  if (num_glyphs > 1)
  {
    code_pair_t range;
    range.code = 1;  // first CID
    range.glyph = num_glyphs - 2;  // nLeft (covers glyphs 1 to num_glyphs-1)
    ranges.push (range);
  }

  if (unlikely (!charset->serialize (c, 2, num_glyphs, ranges)))
  {
    c->pop_discard ();
    return false;
  }

  charset_link = c->pop_pack ();
  return true;
}

/* CFF2 to CFF1 serialization */
namespace CFF {

bool
serialize_cff2_to_cff1 (hb_serialize_context_t *c,
                        OT::cff2_subset_plan &plan,
                        const cff2_top_dict_values_t &cff2_topDict,
                        const OT::cff2::accelerator_subset_t &acc)
{
  TRACE_SERIALIZE (this);

  /*
   * CFF1 Serialization Order (reverse, as HarfBuzz packs from end):
   * 1. CharStrings
   * 2. Private DICs & Local Subrs
   * 3. FDArray
   * 4. FDSelect
   * 5. Charset
   * 6. Global Subrs
   * 7. String INDEX
   * 8. Top DICT INDEX
   * 9. Name INDEX
   * 10. Header
   */

  // 0. Optimize width encoding (for all FDs)
  {
    // Collect widths from hmtx_map
    hb_vector_t<unsigned> widths;
    widths.alloc (plan.num_glyphs);

    for (unsigned gid = 0; gid < plan.num_glyphs; gid++)
    {
      unsigned width = 0;
      if (plan.hmtx_map->has (gid))
        width = plan.hmtx_map->get (gid).first;
      widths.push (width);
    }

    // Optimize defaultWidthX and nominalWidthX
    CFF::optimize_widths (widths, plan.default_width, plan.nominal_width);
  }

  // 1. CharStrings (with widths prepended)
  if (!_serialize_cff1_charstrings (c, plan, plan.default_width, plan.nominal_width))
    return_trace (false);

  // 2. Private DICs & Local Subrs (same as CFF2)
  hb_vector_t<table_info_t> private_dict_infos;
  if (unlikely (!private_dict_infos.resize (plan.subset_fdcount)))
    return_trace (false);

  for (int i = (int)acc.privateDicts.length; --i >= 0;)
  {
    if (plan.fdmap.has (i))
    {
      objidx_t subrs_link = 0;

      if (plan.subset_localsubrs[i].length > 0)
      {
        auto *dest = c->push<CFF1Subrs> ();
        if (likely (dest->serialize (c, plan.subset_localsubrs[i])))
          subrs_link = c->pop_pack (false);
        else
        {
          c->pop_discard ();
          return_trace (false);
        }
      }

      auto *pd = c->push<PrivateDict> ();
      // Use the CFF2 Private DICT serializer which instantiates blends when pinned=true
      cff2_private_dict_op_serializer_t privSzr (plan.desubroutinize, plan.drop_hints, plan.pinned,
                                                  acc.varStore, plan.normalized_coords);
      if (likely (pd->serialize (c, acc.privateDicts[i], privSzr, subrs_link)))
      {
        // Add defaultWidthX and nominalWidthX for CFF1
        str_buff_t width_ops;
        str_encoder_t encoder (width_ops);
        encoder.encode_int (plan.default_width);
        encoder.encode_op (OpCode_defaultWidthX);
        encoder.encode_int (plan.nominal_width);
        encoder.encode_op (OpCode_nominalWidthX);

        if (!encoder.in_error () && c->embed (width_ops.as_bytes ().arrayZ, width_ops.length))
        {
          unsigned fd = plan.fdmap[i];
          private_dict_infos[fd].size = c->length ();
          private_dict_infos[fd].link = c->pop_pack ();
        }
        else
        {
          c->pop_discard ();
          return_trace (false);
        }
      }
      else
      {
        c->pop_discard ();
        return_trace (false);
      }
    }
  }

  // 3. FDArray - serialize CFF2 font dicts as CFF1
  {
    auto *fda = c->push<FDArray<HBUINT16>> ();
    cff_font_dict_op_serializer_t fontSzr;
    auto it =
    + hb_zip (+ hb_iter (acc.fontDicts)
              | hb_filter ([&] (const cff2_font_dict_values_t &_)
                { return plan.fdmap.has (&_ - &acc.fontDicts[0]); }),
              hb_iter (private_dict_infos))
    ;
    // Explicitly specify template parameters: DICTVAL, INFO
    bool success = fda->serialize<cff2_font_dict_values_t, table_info_t> (c, it, fontSzr);
    if (success)
      plan.info.fd_array_link = c->pop_pack (false);
    else
    {
      c->pop_discard ();
      return_trace (false);
    }
  }

  // 4. FDSelect (required in CFF1 CID-keyed fonts)
  // CFF1 requires FDSelect for all CID-keyed fonts, even with just one FD
  // CFF2 makes it optional when there's only one FD
  if (acc.fdSelect != &Null (CFF2FDSelect))
  {
    c->push ();
    if (likely (hb_serialize_cff_fdselect (c, plan.num_glyphs,
                                          *(const FDSelect *)acc.fdSelect,
                                          plan.orig_fdcount,
                                          plan.subset_fdselect_format,
                                          plan.subset_fdselect_size,
                                          plan.subset_fdselect_ranges)))
      plan.info.fd_select.link = c->pop_pack ();
    else
    {
      c->pop_discard ();
      return_trace (false);
    }
  }
  else
  {
    // Create a range-based FDSelect3 mapping all glyphs to FD 0
    // Format: format(1) + nRanges(2) + range(3) + sentinel(2) = 8 bytes
    c->push ();

    // Format byte
    HBUINT8 format;
    format = 3;
    if (unlikely (!c->embed (format)))
    {
      c->pop_discard ();
      return_trace (false);
    }

    // nRanges
    HBUINT16 nRanges;
    nRanges = 1;
    if (unlikely (!c->embed (nRanges)))
    {
      c->pop_discard ();
      return_trace (false);
    }

    // Single range: {first: 0, fd: 0}
    FDSelect3_Range range;
    range.first = 0;
    range.fd = 0;
    if (unlikely (!c->embed (range)))
    {
      c->pop_discard ();
      return_trace (false);
    }

    // Sentinel (number of glyphs)
    HBUINT16 sentinel;
    sentinel = plan.num_glyphs;
    if (unlikely (!c->embed (sentinel)))
    {
      c->pop_discard ();
      return_trace (false);
    }

    plan.info.fd_select.link = c->pop_pack ();
  }

  // 5. Charset (CID charset for identity mapping)
  objidx_t charset_link;
  if (!_serialize_cff1_charset (c, plan.num_glyphs, charset_link))
    return_trace (false);

  // 6. Global Subrs
  {
    auto *dest = c->push<CFF1Subrs> ();
    if (likely (dest->serialize (c, plan.subset_globalsubrs)))
      c->pop_pack (false);
    else
    {
      c->pop_discard ();
      return_trace (false);
    }
  }

  // 7. String INDEX - Add "Adobe" and "Identity" for ROS operator
  {
    const char *adobe_str = "Adobe";
    const char *identity_str = "Identity";
    unsigned adobe_len = 5;  // strlen("Adobe")
    unsigned identity_len = 8;  // strlen("Identity")

    // Build strings array
    hb_vector_t<hb_ubytes_t> strings;
    strings.alloc (2);
    strings.push (hb_ubytes_t ((const unsigned char *) adobe_str, adobe_len));
    strings.push (hb_ubytes_t ((const unsigned char *) identity_str, identity_len));

    // Serialize as CFF INDEX
    auto *dest = c->push<CFF1Index> ();
    if (likely (dest->serialize (c, strings)))
      c->pop_pack (false);
    else
    {
      c->pop_discard ();
      return_trace (false);
    }
  }

  // 8. CFF Header
  OT::cff1 *cff = c->allocate_min<OT::cff1> ();
  if (unlikely (!cff)) return_trace (false);

  /* header */
  cff->version.major = 0x01;
  cff->version.minor = 0x00;
  cff->nameIndex = cff->min_size;
  cff->offSize = 4; /* unused? */

  // 9. Name INDEX (single entry)
  {
    unsigned name_len = strlen (CFF1_DEFAULT_FONT_NAME);

    CFF1Index *idx = c->start_embed<CFF1Index> ();
    if (unlikely (!idx)) return_trace (false);

    if (unlikely (!idx->serialize_header (c, hb_iter (&name_len, 1), name_len)))
      return_trace (false);

    if (unlikely (!c->embed (CFF1_DEFAULT_FONT_NAME, name_len)))
      return_trace (false);
  }

  // 10. Top DICT INDEX
  {
    // Serialize the Top DICT data first
    c->push<TopDict> ();
    cff1_from_cff2_top_dict_op_serializer_t topSzr;

    // Serialize ROS first
    if (unlikely (!topSzr.serialize_ros (c)))
    {
      c->pop_discard ();
      return_trace (false);
    }

    // Serialize FontBBox from head table
    {
      str_buff_t bbox_buff;
      str_encoder_t encoder (bbox_buff);

      encoder.encode_int (plan.head_maxp_info.xMin);
      encoder.encode_int (plan.head_maxp_info.yMin);
      encoder.encode_int (plan.head_maxp_info.xMax);
      encoder.encode_int (plan.head_maxp_info.yMax);
      encoder.encode_op (OpCode_FontBBox);

      if (encoder.in_error () || !c->embed (bbox_buff.as_bytes ().arrayZ, bbox_buff.length))
      {
        c->pop_discard ();
        return_trace (false);
      }
    }

    // Serialize charset operator
    if (charset_link && unlikely (!FontDict::serialize_link4_op (c, OpCode_charset, charset_link, whence_t::Absolute)))
    {
      c->pop_discard ();
      return_trace (false);
    }

    // Serialize FDSelect operator (required for CID-keyed CFF1 fonts)
    if (plan.info.fd_select.link && unlikely (!FontDict::serialize_link4_op (c, OpCode_FDSelect, plan.info.fd_select.link, whence_t::Absolute)))
    {
      c->pop_discard ();
      return_trace (false);
    }

    // Serialize FDArray operator (required for CID-keyed CFF1 fonts)
    if (plan.info.fd_array_link && unlikely (!FontDict::serialize_link4_op (c, OpCode_FDArray, plan.info.fd_array_link, whence_t::Absolute)))
    {
      c->pop_discard ();
      return_trace (false);
    }

    // Serialize other operators from CFF2 TopDict
    for (const auto &opstr : cff2_topDict.values)
    {
      if (unlikely (!topSzr.serialize (c, opstr, plan.info)))
      {
        c->pop_discard ();
        return_trace (false);
      }
    }

    unsigned top_size = c->length ();
    c->pop_pack (false);

    // Serialize INDEX header
    auto *dest = c->start_embed<CFF1Index> ();
    if (unlikely (!dest->serialize_header (c, hb_iter (&top_size, 1), top_size)))
      return_trace (false);
  }

  return_trace (true);
}

} /* namespace CFF */

static bool _serialize_cff2_charstrings (hb_serialize_context_t *c,
			     cff2_subset_plan &plan,
			     const OT::cff2::accelerator_subset_t  &acc)
{
  c->push ();

  unsigned data_size = 0;
  unsigned total_size = CFF2CharStrings::total_size (plan.subset_charstrings, &data_size, plan.min_charstrings_off_size);
  if (unlikely (!c->start_zerocopy (total_size)))
    return false;

  auto *cs = c->start_embed<CFF2CharStrings> ();
  if (unlikely (!cs->serialize (c, plan.subset_charstrings, &data_size, plan.min_charstrings_off_size)))
  {
    c->pop_discard ();
    return false;
  }

  plan.info.char_strings_link = c->pop_pack (false);
  return true;
}

bool
OT::cff2::accelerator_subset_t::serialize (hb_serialize_context_t *c,
					   struct cff2_subset_plan &plan,
					   hb_array_t<int> normalized_coords) const
{
  /* push charstrings onto the object stack first which will ensure it packs as the last
     object in the table. Keeping the chastrings last satisfies the requirements for patching
     via IFTB. If this ordering needs to be changed in the future, charstrings should be left
     at the end whenever HB_SUBSET_FLAGS_ITFB_REQUIREMENTS is enabled. */
  if (!_serialize_cff2_charstrings(c, plan, *this))
    return false;

  /* private dicts & local subrs */
  hb_vector_t<table_info_t>  private_dict_infos;
  if (unlikely (!private_dict_infos.resize (plan.subset_fdcount))) return false;

  for (int i = (int)privateDicts.length; --i >= 0 ;)
  {
    if (plan.fdmap.has (i))
    {
      objidx_t	subrs_link = 0;

      if (plan.subset_localsubrs[i].length > 0)
      {
	auto *dest = c->push <CFF2Subrs> ();
	if (likely (dest->serialize (c, plan.subset_localsubrs[i])))
	  subrs_link = c->pop_pack (false);
	else
	{
	  c->pop_discard ();
	  return false;
	}
      }
      auto *pd = c->push<PrivateDict> ();
      cff2_private_dict_op_serializer_t privSzr (plan.desubroutinize, plan.drop_hints, plan.pinned,
						 varStore, normalized_coords,
						 plan.partial_instancing ? &plan.instancing : nullptr);
      if (likely (pd->serialize (c, privateDicts[i], privSzr, subrs_link)))
      {
	unsigned fd = plan.fdmap[i];
	private_dict_infos[fd].size = c->length ();
	private_dict_infos[fd].link = c->pop_pack ();
      }
      else
      {
	c->pop_discard ();
	return false;
      }
    }
  }

  /* FDSelect */
  if (fdSelect != &Null (CFF2FDSelect))
  {
    c->push ();
    if (likely (hb_serialize_cff_fdselect (c, plan.num_glyphs, *(const FDSelect *)fdSelect,
					   plan.orig_fdcount,
					   plan.subset_fdselect_format, plan.subset_fdselect_size,
					   plan.subset_fdselect_ranges)))
      plan.info.fd_select.link = c->pop_pack ();
    else
    {
      c->pop_discard ();
      return false;
    }
  }

  /* FDArray (FD Index) */
  {
    auto *fda = c->push<CFF2FDArray> ();
    cff_font_dict_op_serializer_t fontSzr;
    auto it =
    + hb_zip (+ hb_iter (fontDicts)
	      | hb_filter ([&] (const cff2_font_dict_values_t &_)
		{ return plan.fdmap.has (&_ - &fontDicts[0]); }),
	      hb_iter (private_dict_infos))
    ;
    if (unlikely (!fda->serialize (c, it, fontSzr)))
    {
      c->pop_discard ();
      return false;
    }
    plan.info.fd_array_link = c->pop_pack (false);
  }

  /* variation store */
  if (varStore != &Null (CFF2ItemVariationStore) &&
      !plan.pinned)
  {
    if (plan.partial_instancing)
    {
      if (plan.instancing.has_variations ())
      {
	c->push ();
	if (unlikely (!plan.instancing.serialize_var_store (c, *plan.axis_tags)))
	{
	  c->pop_discard ();
	  return false;
	}
	plan.info.var_store_link = c->pop_pack (false);
      }
      /* else: the store emptied; the vstore op gets dropped along with it. */
    }
    else
    {
      auto *dest = c->push<CFF2ItemVariationStore> ();
      if (unlikely (!dest->serialize (c, varStore)))
      {
	c->pop_discard ();
	return false;
      }
      plan.info.var_store_link = c->pop_pack (false);
    }
  }

  OT::cff2 *cff2 = c->allocate_min<OT::cff2> ();
  if (unlikely (!cff2)) return false;

  /* header */
  cff2->version.major = 0x02;
  cff2->version.minor = 0x00;
  cff2->topDict = OT::cff2::static_size;

  /* top dict */
  {
    TopDict &dict = cff2 + cff2->topDict;
    cff2_top_dict_op_serializer_t topSzr;
    if (unlikely (!dict.serialize (c, topDict, topSzr, plan.info))) return false;
    cff2->topDictSize = c->head - (const char *)&dict;
  }

  /* global subrs */
  {
    auto *dest = c->start_embed <CFF2Subrs> ();
    return dest->serialize (c, plan.subset_globalsubrs);
  }
}

bool
OT::cff2::accelerator_subset_t::subset (hb_subset_context_t *c) const
{
  cff2_subset_plan cff2_plan;

  if (unlikely (!cff2_plan.create (*this, c->plan))) return false;

  // If instantiating (pinned) and downgrade flag is set, convert to CFF1
  if (cff2_plan.pinned && (c->plan->flags & HB_SUBSET_FLAGS_DOWNGRADE_CFF2))
  {
    // Serialize CFF1 to the subsetter's serializer
    // If we run out of room, returning true will cause subsetter to retry with larger buffer
    bool result = CFF::serialize_cff2_to_cff1 (c->serializer, cff2_plan, topDict, *this);

    if (c->serializer->ran_out_of_room ())
      return true; // Subsetter will retry with larger buffer

    if (result && !c->serializer->in_error ())
    {
      // Success - end serialization to resolve links
      c->serializer->end_serialize ();

      // Copy the serialized CFF1 data and add as CFF table
      hb_blob_t *cff_blob = c->serializer->copy_blob ();
      if (cff_blob)
      {
        c->plan->add_table (HB_TAG('C','F','F',' '), cff_blob);
        hb_blob_destroy (cff_blob);

        // Return false to signal CFF2 table is not needed
        return false;
      }
    }

    // Conversion failed - don't fall back, fail hard for debugging
    return false;
  }

  return serialize (c->serializer, cff2_plan,
		    c->plan->normalized_coords.as_array ());
}

#endif