Note: Strange that FLAG_ALT7 is used but ALT4-ALT6 are not. Same with
R6p.

utf8_nbytes() is only used in dictionary.c. Others are moved to common.c

fb06936c b52d1f3afd and c4d8976fbc broke it.

Use parameters instead of globals. Note the change from sizeof(word_phonemes) to size_word_phonemes.

The file will be organized to have one callable function only. This
should make code structure simpler.

Existing code will be changed to use function parameters instead of
global variables.

Possible problems include too much dependencies with numbers.c.

cmn (Mandarin chinese) has been broken since 4825905.

This fix makes mandarin behave more like Cantonese. Instead of
translating characters, we search for dictionary matches.

The functionality of normal vs Chao tones should be investigated more.
Looks like latin characters as pinyin still uses Chao tones whereas
the characters in cmn_list and cmn_listx do not.

See #1044 for discussion. See also #1028 and #1163.

This is the only reference to the language.

This is the only reference to the language.

Language 'bo' has no data files. Tibetan might have been supported
before but clearly isn't any longer.

ReadClause() is very long and should be shortened for code readability

It's used only in readclause.c

TODO: Rename so the meaning of the function is easier to understand.

./src/espeak-ng -v ar --path=$PWD 5777

would access out of the vowel_stress array, because GetVowelStress, in the
phcode == phonSYLLABIC case, only increments count if stress is 0.  This
code seems fishy, and is not coherent with the loop in SetWordStress.  I
made it similar to the other case above, thus fixing the out-of-bound
access.

This is modifiying the stress in the tn test, I don't know whether this
is expected.

E.g. yue does not have many thousands separators, so the value may be very
large, possibly as much as maxint.

utf8_in2, when working in backward mode, is assuming that we are giving
the address of the last byte of the previous character (see all other
calls to utf8_in2). Otherwise, utf8_in2 returns the size of the current
multibyte character instead of that of the previous multibyte character.

This first reverts "Fix number_buf buffer overflow"
(commit ada93e2db0)

This for loop is apparently actually expected to to skip over NUL
characters.

Fixes #1302

Instead, this limits number processing to 32 digits, as break_numbers does
not support more and would provide bogus result with further digits.

Also fix the signedness of break_numbers so that the 32th bit
actually effectively works.

With emojis, phoneme strings can be very long

TranslateWord2 passes translator2 as tr to TranslateWord which may call
TranslateWord3, SpeakIndividualLetters, TranslateLetter, which was calling
SetTranslator2 again, thus freeing the very tr being used. Make that
latter use another translator.

When there are language switches, when we rewind to the start of the
phoneme list we have to reset the phoneme table back.

This avoids some branching that depends on undefined values, caught by
valgrind in the case e.g. of an emoji substitution that contains a
language switch.

Ref #874

IsLetterGroup looks for heading \0, so the padding also needs to include
an \0.

So valgrind can potentially catch e.g. missing \0.

These don't actually need to survive function exit.