commit 161206455661b697b239d68d72eac48e18e682d5
Author: Daniel Swanson <popcorn.tomato.dude@gmail.com>
Date:   Thu Jul 29 14:00:59 2021 -0500

    transducer mmap sort by input symbol

diff --git a/lttoolbox/Makefile.am b/lttoolbox/Makefile.am
index e943b39..b2875a7 100644
--- a/lttoolbox/Makefile.am
+++ b/lttoolbox/Makefile.am
@@ -2,7 +2,7 @@
 h_sources = alphabet.h att_compiler.h buffer.h compiler.h compression.h  \
             deserialiser.h entry_token.h expander.h fst_processor.h input_file.h lt_locale.h \
             match_exe.h match_node.h match_state.h my_stdio.h node.h \
-            pattern_list.h regexp_compiler.h serialiser.h sorted_vector.h state.h string_utils.h string_writer.h \
+            pattern_list.h regexp_compiler.h serialiser.h sorted_vector.h state.h string_utils.h string_view.h string_writer.h \
             transducer.h trans_exe.h xml_parse_util.h xml_walk_util.h exception.h tmx_compiler.h \
             ustring.h
 cc_sources = alphabet.cc att_compiler.cc compiler.cc compression.cc entry_token.cc \
diff --git a/lttoolbox/transducer.cc b/lttoolbox/transducer.cc
index b281646..7391318 100644
--- a/lttoolbox/transducer.cc
+++ b/lttoolbox/transducer.cc
@@ -741,12 +741,23 @@ Transducer::write_mmap(FILE* out, const Alphabet& alpha)
   }
 
   for (auto& it : transitions) {
+    // we want to make sure the transitions are sorted by input symbol
+    map<int, set<int>> symbols;
     for (auto& it2 : it.second) {
-      auto sym = alpha.decode(it2.first);
-      write_le(out, sym.first); // input symbol
-      write_le(out, sym.second); // output symbol
-      write_le(out, it2.second.first); // destination
-      write_le(out, *reinterpret_cast<uint64_t*>(&it2.second.second)); // weight
+      symbols[alpha.decode(it2.first).first].insert(it2.first);
+    }
+    for (auto& s_in : symbols) {
+      for (auto& s : s_in.second) {
+        auto range = it.second.equal_range(s);
+        for (auto tr = range.first; tr != range.second; ++tr) {
+          auto sym = alpha.decode(tr->first);
+          write_le(out, sym.first); // input symbol
+          write_le(out, sym.second); // output symbol
+          write_le(out, tr->second.first); // destination
+          uint64_t w = *reinterpret_cast<uint64_t*>(&tr->second.second);
+          write_le(out, w); // weight
+        }
+      }
     }
   }
 }