Optimize table layout

Author: eschnett

data/data_generator.jl

@@ -469,12 +469,24 @@ function print_c_data_tables(io, sequences, prop_page_indices, prop_pages, dedup
     end
     print(io, "};\n\n")
 
-    print(io, "static const utf8proc_uint32_t utf8proc_combinations[][2] = {\n")
+    print(io, "static const utf8proc_uint32_t utf8proc_combinations_second[] = {\n")
     for dm0 in sort!(collect(keys(comb_mapping)))
+        print(io, " ");
+        for dm1 in sort!(collect(keys(comb_mapping[dm0])))
+            print(io, " ", dm1, ",")
+        end
+        print(io, "\n");
+    end
+    print(io, "};\n\n")
+
+    print(io, "static const utf8proc_uint32_t utf8proc_combinations_combined[] = {\n")
+    for dm0 in sort!(collect(keys(comb_mapping)))
+        print(io, " ");
         for dm1 in sort!(collect(keys(comb_mapping[dm0])))
             code = comb_mapping[dm0][dm1]
-            print(io, "  {", dm1, ", ", code, "},\n")
+            print(io, " ", code, ",")
         end
+        print(io, "\n");
     end
     print(io, "};\n\n")
 end

utf8proc.c

@@ -687,18 +687,18 @@ UTF8PROC_DLLEXPORT utf8proc_ssize_t utf8proc_normalize_utf32(utf8proc_int32_t *b
         int idx = starter_property->comb_index;
         if (idx < 0x3FF && current_property->comb_issecond) {
           int len = starter_property->comb_length;
-          utf8proc_uint32_t max_second = utf8proc_combinations[idx + len - 1][0];
+          utf8proc_uint32_t max_second = utf8proc_combinations_second[idx + len - 1];
           if (current_char <= max_second) {
             // TODO: binary search? arithmetic search?
             for (int off = 0; off < len; ++off) {
-              utf8proc_uint32_t second = utf8proc_combinations[idx + off][0];
+              utf8proc_uint32_t second = utf8proc_combinations_second[idx + off];
               if (current_char < second) {
                 /* not found */
                 break;
               }
               if (current_char == second) {
                 /* found */
-                utf8proc_uint32_t composition = utf8proc_combinations[idx + off][1];
+                utf8proc_uint32_t composition = utf8proc_combinations_combined[idx + off];
                 *starter = composition;
                 starter_property = NULL;
                 break;

utf8proc.h

@@ -266,10 +266,12 @@ typedef struct utf8proc_property_struct {
    * combining pair, and for most, there are only a handful for
    * possible second characters.
    *
-   * The combining table is stored as `utf8proc_uint32_t
-   * utf8proc_combinations[][2]`. That is, it contains a pair `(second
-   * combining character, combined character)` for every character
-   * that can be a first combining character.
+   * The combining table is stored as sparse matrix in the CSR
+   * (compressed sparse row) format. That is, it is stored as two
+   * arrays, `utf8proc_uint32_t utf8proc_combinations_second[]` and
+   * `utf8proc_uint32_t utf8proc_combinations_combined[]`. These
+   * contain the second combining characters and the combined
+   * character of every combining pair.
    *
    * - `comb_index`: Index into the combining table if this character
    *   is the first character in a combining pair, else 0x3ff