Revised checks for multi-character constants 'xyz'

The previous code for elaborating character constants has a small bug:
the value of a wide character constant consisting of several characters
was normalized to type `int`, while, statically, it has type `wchar_t`.
If `wchar_t` is `unsigned short`, for example, the constant `L'ab'`
would elaborate to 6357090, which is not of type `unsigned short`.

This commit fixes the bug by normalizing wide character constants to type
`wchar_t`, regardless of how many characters they contain.

The previous code was odd in another respect: leading `\0` characters
in multi-character constants were ignored.  Hence, `'\0bcde'` was accepted
while `'abcde'` caused a warning.

This commit implements a more predictable behavior: the number of characters
in a character literal is limited a priori to
      sizeof(type of result) / sizeof(type of each character)

So, for non-wide character constants we can typically have up to 4
characters (sizeof(int) / sizeof(char)), while for wide character
constants we can only have one character.

In effect, multiple-character wide character literals are not supported.
This is allowed by the ISO C99 standard and seems consistent with GCC
and Clang.

Finally, a multi-character constant with too many characters was
reported either as an error (if the computation overflowed the 64-bit
accumulator) or as a warning (otherwise).  Here, we make this an error
in all cases.

GCC and Clang only produce warnings, and truncate the value of the
character constant, but an error feels safer.

0 files changed, 0 insertions, 0 deletions