Identify files by (device, inode), not by path, so that symlinks,
relative paths, case-insensitive paths, or other edge cases
do not result in double includes.
Part of that condition's purpose is to ensure that we read the correct
lexer state; but it's possible now for the fstack to be non-empty
*before* the lexer state is registered, i.e. if there is an error
in the function that registers it.
This causes a NULL pointer deref.
* Implement a '#' prefix for raw identifiers that may alias keywords
* Review comments
* Disallow hashless raw identifiers in interpolations
* Run clang-format
* Implement custom generic tagged union `Either`
This should be more efficient than `std::variant`, while still
keeping runtime safety as it `assert`s when `get`ting values.
* Use `Either` for RPN expressions
* Use `Either` for file stack node data
* Use `Either` for `File` buffer
* Use `Either` for `STRFMT` args
* Use `Either` for RGBLINK symbol values
* Support an equivalent of `std::monostate` for `Either`
* Use `Either` for lexer tokens
* Use `Either` for symbol values
* Use `Either` for lexer mmap/buffer state
Large sizes are more efficient when it's actually buffered,
but most of the time `mmap` is used instead, and the extra size
just slows down allocation of lexer states.
`std::visit` is (arguably) cleaner code, but older versions of gcc
and clang (not very old; the ones packaged with Ubuntu 22.04 LTS)
compile them as tables of function pointers, instead of efficient
jump tables.
- Since we have style rules to include foo.hpp at the top of its
corresponding foo.cpp, this takes any headers included by foo.hpp
as being also guaranteed for foo.cpp.
- Use C-style <foo.h> instead of <cfoo>, since the latter only
guarantees putting symbols in the `std` namespace, which we are
not using for C functions (e.g. `printf` not `std::printf`).
- Remove now-unused `__PRETTY_FUNCTION__` reporting
