+ *
+ * Changes from format version 2 to format version 3 (ICU 60) ------------------
+ *
+ * - norm16 bit 0 indicates hasCompBoundaryAfter(),
+ * except that for contiguous composition (FCC) the tccc must be checked as well.
+ * Data indexes and ccc values are shifted left by one (OFFSET_SHIFT).
+ * Thresholds like minNoNo are tested before shifting.
+ *
+ * - Algorithmic mapping deltas are shifted left by two more bits (total DELTA_SHIFT),
+ * to make room for two bits (three values) indicating whether the tccc is 0, 1, or greater.
+ * See DELTA_TCCC_MASK etc.
+ * This helps with fetching tccc/FCD values and FCC hasCompBoundaryAfter().
+ * minMaybeYes is 8-aligned so that the DELTA_TCCC_MASK bits can be tested directly.
+ *
+ * - Algorithmic mappings are only used for mapping to "comp yes and ccc=0" characters,
+ * and ASCII characters are mapped algorithmically only to other ASCII characters.
+ * This helps with hasCompBoundaryBefore() and compose() fast paths.
+ * It is never necessary any more to loop for algorithmic mappings.
+ *
+ * - Addition of indexes[IX_MIN_NO_NO_COMP_BOUNDARY_BEFORE],
+ * indexes[IX_MIN_NO_NO_COMP_NO_MAYBE_CC], and indexes[IX_MIN_NO_NO_EMPTY],
+ * and separation of the noNo extraData into distinct ranges.
+ * With this, the noNo norm16 value indicates whether the mapping is
+ * compose-normalized, not normalized but hasCompBoundaryBefore(),
+ * not even that, or maps to an empty string.
+ * hasCompBoundaryBefore() can be determined solely from the norm16 value.
+ *
+ * - The norm16 value for Hangul LVT is now different from that for Hangul LV,
+ * so that hasCompBoundaryAfter() need not check for the syllable type.
+ * For Hangul LV, minYesNo continues to be used (no comp-boundary-after).
+ * For Hangul LVT, minYesNoMappingsOnly|HAS_COMP_BOUNDARY_AFTER is used.
+ * The extraData units at these indexes are set to firstUnit=2 and firstUnit=3, respectively,
+ * to simplify some code.
+ *
+ * - The extraData firstUnit bit 5 is no longer necessary
+ * (norm16 bit 0 used instead of firstUnit MAPPING_NO_COMP_BOUNDARY_AFTER),
+ * is reserved again, and always set to 0.
+ *
+ * - Addition of indexes[IX_MIN_LCCC_CP], the first code point where lccc!=0.
+ * This used to be hardcoded to U+0300, but in data like NFKC_Casefold it is lower:
+ * U+00AD Soft Hyphen maps to an empty string,
+ * which is artificially assigned "worst case" values lccc=1 and tccc=255.
+ *
+ * - A mapping to an empty string has explicit lccc=1 and tccc=255 values.
+ *
+ * Changes from format version 3 to format version 4 (ICU 63) ------------------
+ *
+ * Switched from UTrie2 to UCPTrie/CodePointTrie.
+ *
+ * The new trie no longer stores different values for surrogate code *units* vs.
+ * surrogate code *points*.
+ * Lead surrogates still have values for optimized UTF-16 string processing.
+ * When looking up code point properties, the code now checks for lead surrogates and
+ * treats them as inert.
+ *
+ * gennorm2 now has to reject mappings for surrogate code points.
+ * UTS #46 maps unpaired surrogates to U+FFFD in code rather than via its
+ * custom normalization data file.