nushell/crates/nu-command/src/input_handler.rs
Stefan Holderbach a52386e837
Box ShellError in Value::Error (#8375)
# Description

Our `ShellError` at the moment has a `std::mem::size_of<ShellError>` of
136 bytes (on AMD64). As a result `Value` directly storing the struct
also required 136 bytes (thanks to alignment requirements).

This change stores the `Value::Error` `ShellError` on the heap.

Pro:
- Value now needs just 80 bytes
- Should be 1 cacheline less (still at least 2 cachelines)

Con:
- More small heap allocations when dealing with `Value::Error`
  - More heap fragmentation
  - Potential for additional required memcopies

# Further code changes

Includes a small refactor of `try` due to a type mismatch in its large
match.

# User-Facing Changes

None for regular users.

Plugin authors may have to update their matches on `Value` if they use
`nu-protocol`

Needs benchmarking to see if there is a benefit in real world workloads.
**Update** small improvements in runtime for workloads with high volume
of values. Significant reduction in maximum resident set size, when many
values are held in memory.

# Tests + Formatting
2023-03-12 09:57:27 +01:00

91 lines
2.8 KiB
Rust

use nu_protocol::ast::CellPath;
use nu_protocol::{PipelineData, ShellError, Span, Value};
use std::sync::atomic::AtomicBool;
use std::sync::Arc;
pub trait CmdArgument {
fn take_cell_paths(&mut self) -> Option<Vec<CellPath>>;
}
/// Arguments with only cell_path.
///
/// If commands is going to use `operate` function, and it only required optional cell_paths
/// Using this to simplify code.
pub struct CellPathOnlyArgs {
cell_paths: Option<Vec<CellPath>>,
}
impl CmdArgument for CellPathOnlyArgs {
fn take_cell_paths(&mut self) -> Option<Vec<CellPath>> {
self.cell_paths.take()
}
}
impl From<Vec<CellPath>> for CellPathOnlyArgs {
fn from(cell_paths: Vec<CellPath>) -> Self {
Self {
cell_paths: (!cell_paths.is_empty()).then_some(cell_paths),
}
}
}
/// A simple wrapper for `PipelineData::map` method.
///
/// In detail, for each elements, invoking relative `cmd` with `arg`.
///
/// If `arg` tell us that its cell path is not None, only map over data under these columns.
/// Else it will apply each column inside a table.
///
/// The validation of input element should be handle by `cmd` itself.
pub fn operate<C, A>(
cmd: C,
mut arg: A,
input: PipelineData,
span: Span,
ctrlc: Option<Arc<AtomicBool>>,
) -> Result<PipelineData, ShellError>
where
A: CmdArgument + Send + Sync + 'static,
C: Fn(&Value, &A, Span) -> Value + Send + Sync + 'static + Clone + Copy,
{
match arg.take_cell_paths() {
None => input.map(
move |v| {
match v {
// Propagate errors inside the input
Value::Error { .. } => v,
_ => cmd(&v, &arg, span),
}
},
ctrlc,
),
Some(column_paths) => {
let arg = Arc::new(arg);
input.map(
move |mut v| {
for path in &column_paths {
let opt = arg.clone();
let r = v.update_cell_path(
&path.members,
Box::new(move |old| {
match old {
// Propagate errors inside the input
Value::Error { .. } => old.clone(),
_ => cmd(old, &opt, span),
}
}),
);
if let Err(error) = r {
return Value::Error {
error: Box::new(error),
};
}
}
v
},
ctrlc,
)
}
}
}