Specifically refactor u_get_undo_file_name which will be modified to
automatically create undo directory and replace `char_u` with `char` in some of
the related functions.
a8e18d9 introduced an off-by-one error that caused the last character of
&directory to be ignored. This commit is a straightforward fix for that error.
fixes#3519
Begin refactoring getcmdline() into a state that can be managed by the
`state_enter()`:
- Move local variables into a local CommandLineState structure
- Fix code style in the entire function.
- Create `TerminalState` structure containing data used in terminal mode
- Extract `terminal_execute` from `terminal_enter` and use it with
`state_enter`.
- `insert_handle_key`: Contains the big insert mode switch statement.
- `insert_do_complete`: Code that used to be in the `docomplete` label.
- `insert_do_cindent`: Code that used to be in the `force_cindent` label.
Also move some code after the switch statement into the beginning of
`insert_check`.
Refactor insert mode to use `state_enter` as an event loop:
- Move o_lnum(static variable) outside function
- Move code before the insert mode loop into `insert_enter`
- Move code before `safe_vgetc()` call into `insert_check`
- Move code after `safe_vgetc()` call into `insert_execute`
- Remove doESCkey label and handle insert mode repeating in the `insert_enter`
function
- Remove do_intr label(this is not the place for platform-specific interrupt
charts)
Begin refactoring edit() into a state that can be managed by the `state_enter()`:
- Move local variables into a local InsertState structure
- Fix code style in the entire function.
From a very high level point of view, Vim/Nvim can be described as state
machines following these instructions in a loop:
- Read user input
- Peform some action. The action is determined by the current state and can
switch states.
- Possibly display some feedback to the user.
This is not immediately visible because these instructions spread across dozens
of nested loops and function calls, making it very hard to modify the state
machine(to support more event types, for example).
So far, the approach Nvim has taken to allow more events is this:
- At the very core function that blocks for input, poll for arbitrary events.
- If the event received from the OS is user input, just return it normally to
the callers.
- If the event is not direct result of user input(possibly a vimscript function
call coming from a msgpack-rpc socket or a job control callback), return a
special key code(`K_EVENT`) that is handled by callers where it is safer to
perform arbitrary actions.
One problem with this approach is that the `K_EVENT` signal is being sent across
multiple states that may be unaware of it. This was partially fixed with the
`input_enable_events`/`input_disable_events` functions, which were added as a
mechanism that the upper layers can use to tell the core input functions that it
is ready to accept `K_EVENT`.
Another problem is that the mapping engine is implemented in getchar.c
which is called from every state, but the mapping engine is not aware of
`K_EVENT` so events can break mappings.
While it is theoretically possible to modify getchar.c to make it aware of
`K_EVENT`, this commit fixes the problem with a different approach: Model Nvim
as a pushdown automaton(https://en.wikipedia.org/wiki/Pushdown_automaton). This
design has many advantages which include:
- Decoupling the event loop from the states reponsible for handling events.
- Better control of state transition with less dependency on global variable
hacks(eg: 'restart_edit' global variable).
- Easier removal of global variables and function splitting. That is because
many variables are for state-specific information, and probably ended up being
global to simplify communication between functions, which we fix by storing
state-specific information in specialized structures.
The final goal is to let Nvim have a single top-level event loop represented by
the following pseudo-code:
```
while not quitting
let event = read_event
current_state(event)
update_screen()
```
This closely mirrors the state machine description above and makes it easier to
understand, extend and debug the program.
Note that while the pseudo code suggests an explicit stack of states that
doesn't rely on return addresses(as suggested by the principles of
automata-based programming:
https://en.wikipedia.org/wiki/Automata-based_programming), for now we'll use the
call stack as a structure to manage state transitioning as it would be very
difficult to refactor Nvim to use an explicit stack of states, and the benefits
would be small.
While this change may seem like an endless amount of work, it is possible to
do it incrementally as was shown in the previous commits. The general procedure
is:
1- Find a blocking `vgetc()`(or derivatives) call. This call represents an
implicit state of the program.
2- Split the code before and after the `vgetc()` call into functions that match
the signature of `state_check_callback` and `state_execute_callback.
Only `state_execute_callback` is required.
3- Create a `VimState` "subclass" and a initializer function that sets the
function pointers and performs any other required initialization steps. If
the state has no local variables, just use `VimState` without subclassing.
4- Instead of calling the original function containing the `vgetc()`,
initialize a stack-allocated `VimState` subclass, then call `state_enter` to
begin processing events in the state.
5- The check/execute callbacks can return 1 to continue normally, 0 to break the
loop or -1 to skip to the next iteration. These callbacks contain code that
execute before and after the old `vgetc()` call.
The functions created in step 2 may contain other `vgetc()` calls. These
represent implicit sub-states of the current state, but it is fine to remove
them later in smaller steps since we didn't break compatibility with existing
code.
This makes it impossible for K_EVENT to interfere with mappings, but it also
disables processing of events while in the middle of a mapping (Though this will
be fixed later as this refactoring progresses).
`may_sync_undo` is now called when K_EVENT is received. This is necessary to
correctly update undo entry lists before executing some action.
`normal_prepare` is now called by `normal_check` before returning 1(to
continue).
Also remove `input_{enable,disable}_events` calls from `normal_cmd`, which only
exists now as a compatibility function to run normal commands with keys inserted
into the typeahead buffer(We don't want to process events in these cases
anyway).
