For now, we are focusing only on iterating over the argument ids and not
the values.
This provides a building block for more obscure use cases like iterating
over argument values, in order. We are not providing it out of the box
at the moment both to not overly incentize a less common case, because
it would abstract away a performance hit, and because we want to let
people experiment with this and if a common path emerges we can consider
it then if there is enough users.
Fixes#1206
Now that `Id` is public, we can have `ArgMatches` report them. If we
have to choose one behavior, this is more universal. The user can still
look up the values, this works with groups whose args have different
types, and this allows people to make decisions off of it when otherwise
there isn't enogh information.
Fixes#2317Fixes#3748
This is a step towards #1041
- `ArgGroup` no longer takes a lifetime
- One less field type needs a lifetime
For now, we are using a more brute force type (`String`) so we can
establish performance base lines. I was torn on whether to use `&str`
everywhere or make an `IdRef`. The latter would add a lot of noise that
I'm concerned about, so i left it simple for now. `IdRef` would help to
communicate the types involved though.
Speaking of communicating types, I'm also torn on whether we should use
`Id` for all strings or if we should have `Id`, `Name`, etc types to
avoid people mixing and matching.
This added 18.7 KB.
Compared to `HEAD~` on `06_rustup`:
- build: 6.23us -> 7.41us
- parse: 8.17us -> 9.36us
- parse_sc: 7.65us -> 9.29us
This is a part of #2870 and is prep for #1041
Oddly enough, this dropped the binary size by 200 Bytes
Compared to `HEAD~` on `06_rustup`:
- build: 6.21us -> 6.23us
- parse: 7.55us -> 8.17us
- parse_sc: 7.95us -> 7.65us
This dropped 17KB
Again, performance shouldn't be too bad as the total number of argument
id's passed in by the user shouldn't be huge, with the upper end being
5-15 except for in extreme cases like rustc accepting arguments from
cargo via a file.
This dropped `.text` by 14KB
Anything in debug asserts or help/usage output doesn't matter for
performance but I wouldn't be surprised if this was comparable since the
container sizes we are talking about are relatively small.
Documenting the existing behavior is challenging which suggests it can
cause user confusion. So long as its not too hard to explicitly
specify actions, we should just do it.
Fixes#4057
Before we introduced actions, it required specific setups to engage with
claps version and help printing. With actions making that more
explicit, we don't get as much benefit from our multiple, obscure, ways
of users customizing help
Before
- Modify existing help or version with `mut_arg` which would
automatically be pushed down the command tree like `global(true)`
- Create an new help or version and have it treated as if it was the
built-in on (I think)
- Use the same flags as built-in and have the built-in flags
automatically disabled
- Users could explicitly disable the built-in functionality and do what
they want
Now
- `mut_arg` no longer works as we define help and version flags at the
end
- If someone defines a flag that overlaps with the built-ins by id,
long, or short, a debug assert will tell them to explicitly disable
the built-in
- Any customization has to be done by a user providing their own. To
propagate through the command tree, they need to set `global(true)`.
Benefits
- Hopefully, this makes it less confusing on how to override help
behavior. Someone creates an arg and we then tell them how to disable
the built-in
- This greatly simplifies the arg handling by pushing more
responsibility onto the developer in what are hopefully just corner
cases
- This removes about 1Kb from .text
Fixes#3405Fixes#4033
Previously the Arg id was set with the "name" attribute. This allows use
of an "id" attribute to match the underlying struct.
A side effect of this is that the "id" attribute may also be used on
Commands. This isn't desired, but given the current architecture of the
attribute parser, it's hard to avoid.
Fixes: #3785
Naming is hard. I found I was writing new code without the `s` so that
suggests the name was wrong. But we renamed `number_of_values` to
`num_args`, so should this be `ArgRange`?
TakesValue helps because it provides a way for a lot of settings to say
a value is needed without specifying how many. Multiple values didn't
have enough call sites to make this worthwhile.
This is a part of #2688
multiple_values is now just book keeping for the builder, instead people
should look to actions and `num_args`.
The meaning for it was a little weird anyways.
The only time it won't be initialized is before `_build`. This is possible because
of #4027
I wish I could just put the `expect` inside the call but I'm worried
about allowing people to build stuff on top of clap.
As we move people off takes_value/multiple_values, this will provide
something to check.
This was previously blocked on other issues related to flag handling
(#4023)
In clap v3, `require_value_delimiter` activated an alternative parse
mode where
- `multiple_values` meant "multiple values within a single arg"
- `number_of_values` having no parse impact, only validation impact
- `value_names` being delimited values
For unbounded `number_of_values`, this is exactly what `value_delimiter`
provides. The only value is if someone wanted `value_name` to be
`<file1>,<file2>,...` which can be useful and we might look into adding
back in.
Alternatively, this could be used for cases like key-value pairs but
that has issues like not allowing the delimiter in the value which might
be ok in some cases but not others. We already instead document that
people should instead use `ValueParser` for this case.
In removing this, we remove points of confusion at how the different
multiple values and delimited value calls interact with each other. I
know I would set `require_value_delimiter(true).multiple_values(true)`
when it turns out all I needed was `value_delimiter(',')`.
This also reduces the API surface area which makes it easier to discover
what features we do provide.
While this isn't big, this is also yet another small step towards
reducing binary size and compile times.
This will allow `num_args(0..=1).value_delimiter(',')` to work properly.
This hacks in support for `require_value_delimiter` until we can remove
it.
This no longer recognzes value terminators in delimited lists.
It looks like there is a bug with recognizing value terminators in
positionals arguments. We'll need to dig into that more.
This reduces ambiguity in how the different "multiple" parts of the API
interact and lowrs the amount of API surface area users have to dig
through to use clap.
For now, this is only a matter of cleaning up the public API. Cleaning
up the implementation is the next step.
When suggesting required arguments, we wanted to avoid an argument
showing up in both a group and by itself but we didn't correctly
calculate that, causing no required arguments to show up at times.
Now, we all use the same pool of information for doing the calculations.
This was the type of cleanup that I expected it to drop our binary size
but this added 1k to our .text. Strange.
Fixes#4004
In the short term, this just provides a back door to custom actions.
Longer term, we can explore a `SetConst` action that relies on this
behavior. Really, `SetTrue` and `SetFalse` are shortcuts for such an
action but shortcuts can be helpful for usability.
Apparently, this also reduced `.text` size by 1k
Rendering of usage is not in a critical path, so should be more worried
about binary size than performance. That only leaves avoiding coloring
spaces. That shouldn't be a problem, so let's make the binary smaller.
With `number_of_values` being per-occurrence now, its doesn't make sense
for `number_of_values(0)` to set `takes_value(true)` or for
`number_of_values(1)` to set `multiple_values(true)`.
In addition, an assert is made if the user works around this
This both simplifies the code and the model we present to the user,
making more sense.
There is room for further exploration of tying flag actions into this.
This changes the default type as well to encourage preserving the full
information for shelling out. If people need UTF-8, then they can
change the value parser.
Fixes#3733
When checking into binary size, I noticed that the `git` example is a
lot larger than v3. `git bisect` narrowed it down to
11076a5c70 which doesn't make sense. I
did noticed we could remove some bloat from monomorphization.
Overall for `cargo-example, we've dropped about 47 KiB.
Previous behavior:
- They'd be sorted by default
- They'd derive display order if `DeriveDisplayOrder` was set
- This could be set recursively
- The initial display order value for subcommands was 0
New behavior:
- Sorted order is derived by default
- Sorting is turned on by `cmd.next_display_order(None)`
- This is not recursive, it must be set on each level
- The display order incrementing is mixed with arguments
- This does make it slightly more difficult to predict
We had some tests for this but not sufficient obviously. The problem is
we were tweaking the positional argument counter when processing flags
and not just positional arguments. Delaying it until after flags seems
to fix this.
Fixes#3959
A couple of things happened when preparing to release 3.0
- We needed derive documentation
- I had liked how serde handled theres
- I had bad experiences finding things in structopt's documentation
- The examples were broken and we needed tests
- The examples seemed to follow a pattern of having tutorial content and
cookbook content
- We had been getting bug reports from people looking at master and
thinking they were looking at what is currently released
- We had gotten feedback to keep down the number of places that
documentation was located
From this, we went with a mix of docs.rs and github
- We kept the number of content locations at 2 rather than 3 by not
having an external site like serde
- We rewrote the examples into explicit tutorials and cookbooks to align
with the 4 styles of documentation
- We could test our examples by running `console` code blocks with
trycmd
- Documentation was versioned and the README pointed to the last release
This had downsides
- The tutorials didn't have the code inlined
- Users still had a hard time finding and navigating between the
different forms of documentation
- In practice, we were less likely to cross-link between the different
types of documentation
Moving to docs.rs would offer a lot of benefits, even if it is only
designed for Rust-reference documentation and isn't good for Rust derive
reference documentation, tutorials, cookbooks, etc. The big problem was
keeping the examples tested to keep maintenance costs down. Maybe its
just me but its easy to overlook
- You can pull documentation from a file using `#[doc = "path"]`
- Repeated doc attributes get concatenated rather than first or last
writer winning
Remember these when specifically thinking about Rust documentation made
me realize that we could get everything into docs.rs.
When doing this
- Tutorial code got brought in as was one of the aims
- We needed to split the lib documentation and the README to have all of
the linking work. This allowed us to specialize them according to
their rule (user vs contributor)
- We needed to avoid users getting caught up in making a decision
between Derive and Builder APIs so we put the focus on the derive API
with links to the FAQ to help users decide when to use one or the
other.
- Improved cross-referencing between different parts of the
documentation
- Limited inline comments were added to example code
- Introductory example code intentionally does not have teaching
comments in it as its meant to give a flavor or sense of things and
not meant to teach on its own.
This is a first attempt. There will be a lot of room for further
improvement. Current know downsides:
- Content source is more split up for the tutorials
This hopefully addresses #3189
