fish-shell/src/ast.rs

4041 lines
134 KiB
Rust
Raw Normal View History

Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
/*!
* This defines the fish abstract syntax tree.
* The fish ast is a tree data structure. The nodes of the tree
* are divided into three categories:
*
* - leaf nodes refer to a range of source, and have no child nodes.
* - branch nodes have ONLY child nodes, and no other fields.
* - list nodes contain a list of some other node type (branch or leaf).
*
* Most clients will be interested in visiting the nodes of an ast.
*/
use crate::common::{unescape_string, UnescapeStringStyle};
use crate::flog::{FLOG, FLOGF};
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
use crate::parse_constants::{
token_type_user_presentable_description, ParseError, ParseErrorCode, ParseErrorList,
ParseKeyword, ParseTokenType, ParseTreeFlags, SourceRange, StatementDecoration,
ERROR_BAD_COMMAND_ASSIGN_ERR_MSG, INVALID_PIPELINE_CMD_ERR_MSG, SOURCE_OFFSET_INVALID,
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
};
use crate::parse_tree::ParseToken;
#[cfg(test)]
use crate::tests::prelude::*;
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
use crate::tokenizer::{
variable_assignment_equals_pos, TokFlags, TokenType, Tokenizer, TokenizerError,
TOK_ACCEPT_UNFINISHED, TOK_CONTINUE_AFTER_ERROR, TOK_SHOW_COMMENTS,
};
2023-08-09 06:16:04 +08:00
use crate::wchar::prelude::*;
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
use std::ops::{ControlFlow, Index, IndexMut};
/**
* A NodeVisitor is something which can visit an AST node.
*
* To visit a node's fields, use the node's accept() function:
* let mut v = MyNodeVisitor{};
* node.accept(&mut v);
*/
pub trait NodeVisitor<'a> {
fn visit(&mut self, node: &'a dyn Node);
}
pub trait Acceptor {
fn accept<'a>(&'a self, visitor: &mut dyn NodeVisitor<'a>, reversed: bool);
}
impl<T: Acceptor> Acceptor for Option<T> {
fn accept<'a>(&'a self, visitor: &mut dyn NodeVisitor<'a>, reversed: bool) {
match self {
Some(node) => node.accept(visitor, reversed),
None => (),
}
}
}
pub struct MissingEndError {
allowed_keywords: &'static [ParseKeyword],
token: ParseToken,
}
pub type VisitResult = ControlFlow<MissingEndError>;
trait NodeVisitorMut {
/// will_visit (did_visit) is called before (after) a node's fields are visited.
fn will_visit_fields_of(&mut self, node: &mut dyn NodeMut);
fn visit_mut(&mut self, node: &mut dyn NodeMut) -> VisitResult;
fn did_visit_fields_of<'a>(&'a mut self, node: &'a dyn NodeMut, flow: VisitResult);
fn visit_argument_or_redirection(
&mut self,
_node: &mut Box<ArgumentOrRedirectionVariant>,
) -> VisitResult;
fn visit_block_statement_header(
&mut self,
_node: &mut Box<BlockStatementHeaderVariant>,
) -> VisitResult;
fn visit_statement(&mut self, _node: &mut Box<StatementVariant>) -> VisitResult;
fn visit_decorated_statement_decorator(
&mut self,
_node: &mut Option<DecoratedStatementDecorator>,
);
fn visit_job_conjunction_decorator(&mut self, _node: &mut Option<JobConjunctionDecorator>);
fn visit_else_clause(&mut self, _node: &mut Option<ElseClause>);
fn visit_semi_nl(&mut self, _node: &mut Option<SemiNl>);
fn visit_time(&mut self, _node: &mut Option<KeywordTime>);
fn visit_token_background(&mut self, _node: &mut Option<TokenBackground>);
}
trait AcceptorMut {
fn accept_mut(&mut self, visitor: &mut dyn NodeVisitorMut, reversed: bool);
}
impl<T: AcceptorMut> AcceptorMut for Option<T> {
fn accept_mut(&mut self, visitor: &mut dyn NodeVisitorMut, reversed: bool) {
match self {
Some(node) => node.accept_mut(visitor, reversed),
None => (),
}
}
}
/// Node is the base trait of all AST nodes.
pub trait Node: Acceptor + ConcreteNode + std::fmt::Debug {
/// The parent node, or null if this is root.
fn parent(&self) -> Option<&dyn Node>;
/// The type of this node.
fn typ(&self) -> Type;
/// The category of this node.
fn category(&self) -> Category;
/// \return a helpful string description of this node.
fn describe(&self) -> WString {
let mut res = ast_type_to_string(self.typ()).to_owned();
if let Some(n) = self.as_token() {
let token_type = n.token_type().to_wstr();
sprintf!(=> &mut res, " '%ls'", token_type);
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
} else if let Some(n) = self.as_keyword() {
let keyword = n.keyword().to_wstr();
sprintf!(=> &mut res, " '%ls'", keyword);
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
res
}
/// \return the source range for this node, or none if unsourced.
/// This may return none if the parse was incomplete or had an error.
fn try_source_range(&self) -> Option<SourceRange>;
/// \return the source range for this node, or an empty range {0, 0} if unsourced.
fn source_range(&self) -> SourceRange {
self.try_source_range().unwrap_or(SourceRange::new(0, 0))
}
/// \return the source code for this node, or none if unsourced.
fn try_source<'s>(&self, orig: &'s wstr) -> Option<&'s wstr> {
self.try_source_range().map(|r| &orig[r.start()..r.end()])
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
/// \return the source code for this node, or an empty string if unsourced.
fn source<'s>(&self, orig: &'s wstr) -> &'s wstr {
self.try_source(orig).unwrap_or_default()
}
// The address of the object, for comparison.
fn as_ptr(&self) -> *const ();
fn pointer_eq(&self, rhs: &dyn Node) -> bool {
std::ptr::eq(self.as_ptr(), rhs.as_ptr())
}
fn as_node(&self) -> &dyn Node;
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
/// NodeMut is a mutable node.
trait NodeMut: Node + AcceptorMut + ConcreteNodeMut {}
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
pub trait ConcreteNode {
// Cast to any sub-trait.
fn as_leaf(&self) -> Option<&dyn Leaf> {
None
}
fn as_keyword(&self) -> Option<&dyn Keyword> {
None
}
fn as_token(&self) -> Option<&dyn Token> {
None
}
// Cast to any node type.
fn as_redirection(&self) -> Option<&Redirection> {
None
}
fn as_variable_assignment(&self) -> Option<&VariableAssignment> {
None
}
fn as_variable_assignment_list(&self) -> Option<&VariableAssignmentList> {
None
}
fn as_argument_or_redirection(&self) -> Option<&ArgumentOrRedirection> {
None
}
fn as_argument_or_redirection_list(&self) -> Option<&ArgumentOrRedirectionList> {
None
}
fn as_statement(&self) -> Option<&Statement> {
None
}
fn as_job_pipeline(&self) -> Option<&JobPipeline> {
None
}
fn as_job_conjunction(&self) -> Option<&JobConjunction> {
None
}
fn as_for_header(&self) -> Option<&ForHeader> {
None
}
fn as_while_header(&self) -> Option<&WhileHeader> {
None
}
fn as_function_header(&self) -> Option<&FunctionHeader> {
None
}
fn as_begin_header(&self) -> Option<&BeginHeader> {
None
}
fn as_block_statement(&self) -> Option<&BlockStatement> {
None
}
fn as_if_clause(&self) -> Option<&IfClause> {
None
}
fn as_elseif_clause(&self) -> Option<&ElseifClause> {
None
}
fn as_elseif_clause_list(&self) -> Option<&ElseifClauseList> {
None
}
fn as_else_clause(&self) -> Option<&ElseClause> {
None
}
fn as_if_statement(&self) -> Option<&IfStatement> {
None
}
fn as_case_item(&self) -> Option<&CaseItem> {
None
}
fn as_switch_statement(&self) -> Option<&SwitchStatement> {
None
}
fn as_decorated_statement(&self) -> Option<&DecoratedStatement> {
None
}
fn as_not_statement(&self) -> Option<&NotStatement> {
None
}
fn as_job_continuation(&self) -> Option<&JobContinuation> {
None
}
fn as_job_continuation_list(&self) -> Option<&JobContinuationList> {
None
}
fn as_job_conjunction_continuation(&self) -> Option<&JobConjunctionContinuation> {
None
}
fn as_andor_job(&self) -> Option<&AndorJob> {
None
}
fn as_andor_job_list(&self) -> Option<&AndorJobList> {
None
}
fn as_freestanding_argument_list(&self) -> Option<&FreestandingArgumentList> {
None
}
fn as_job_conjunction_continuation_list(&self) -> Option<&JobConjunctionContinuationList> {
None
}
fn as_maybe_newlines(&self) -> Option<&MaybeNewlines> {
None
}
fn as_case_item_list(&self) -> Option<&CaseItemList> {
None
}
fn as_argument(&self) -> Option<&Argument> {
None
}
fn as_argument_list(&self) -> Option<&ArgumentList> {
None
}
fn as_job_list(&self) -> Option<&JobList> {
None
}
}
trait ConcreteNodeMut {
// Cast to any sub-trait.
fn as_mut_leaf(&mut self) -> Option<&mut dyn Leaf> {
None
}
fn as_mut_keyword(&mut self) -> Option<&mut dyn Keyword> {
None
}
fn as_mut_token(&mut self) -> Option<&mut dyn Token> {
None
}
// Cast to any node type.
fn as_mut_redirection(&mut self) -> Option<&mut Redirection> {
None
}
fn as_mut_variable_assignment(&mut self) -> Option<&mut VariableAssignment> {
None
}
fn as_mut_variable_assignment_list(&mut self) -> Option<&mut VariableAssignmentList> {
None
}
fn as_mut_argument_or_redirection(&mut self) -> Option<&mut ArgumentOrRedirection> {
None
}
fn as_mut_argument_or_redirection_list(&mut self) -> Option<&mut ArgumentOrRedirectionList> {
None
}
fn as_mut_statement(&mut self) -> Option<&mut Statement> {
None
}
fn as_mut_job_pipeline(&mut self) -> Option<&mut JobPipeline> {
None
}
fn as_mut_job_conjunction(&mut self) -> Option<&mut JobConjunction> {
None
}
fn as_mut_for_header(&mut self) -> Option<&mut ForHeader> {
None
}
fn as_mut_while_header(&mut self) -> Option<&mut WhileHeader> {
None
}
fn as_mut_function_header(&mut self) -> Option<&mut FunctionHeader> {
None
}
fn as_mut_begin_header(&mut self) -> Option<&mut BeginHeader> {
None
}
fn as_mut_block_statement(&mut self) -> Option<&mut BlockStatement> {
None
}
fn as_mut_if_clause(&mut self) -> Option<&mut IfClause> {
None
}
fn as_mut_elseif_clause(&mut self) -> Option<&mut ElseifClause> {
None
}
fn as_mut_elseif_clause_list(&mut self) -> Option<&mut ElseifClauseList> {
None
}
fn as_mut_else_clause(&mut self) -> Option<&mut ElseClause> {
None
}
fn as_mut_if_statement(&mut self) -> Option<&mut IfStatement> {
None
}
fn as_mut_case_item(&mut self) -> Option<&mut CaseItem> {
None
}
fn as_mut_switch_statement(&mut self) -> Option<&mut SwitchStatement> {
None
}
fn as_mut_decorated_statement(&mut self) -> Option<&mut DecoratedStatement> {
None
}
fn as_mut_not_statement(&mut self) -> Option<&mut NotStatement> {
None
}
fn as_mut_job_continuation(&mut self) -> Option<&mut JobContinuation> {
None
}
fn as_mut_job_continuation_list(&mut self) -> Option<&mut JobContinuationList> {
None
}
fn as_mut_job_conjunction_continuation(&mut self) -> Option<&mut JobConjunctionContinuation> {
None
}
fn as_mut_andor_job(&mut self) -> Option<&mut AndorJob> {
None
}
fn as_mut_andor_job_list(&mut self) -> Option<&mut AndorJobList> {
None
}
fn as_mut_freestanding_argument_list(&mut self) -> Option<&mut FreestandingArgumentList> {
None
}
fn as_mut_job_conjunction_continuation_list(
&mut self,
) -> Option<&mut JobConjunctionContinuationList> {
None
}
fn as_mut_maybe_newlines(&mut self) -> Option<&mut MaybeNewlines> {
None
}
fn as_mut_case_item_list(&mut self) -> Option<&mut CaseItemList> {
None
}
fn as_mut_argument(&mut self) -> Option<&mut Argument> {
None
}
fn as_mut_argument_list(&mut self) -> Option<&mut ArgumentList> {
None
}
fn as_mut_job_list(&mut self) -> Option<&mut JobList> {
None
}
}
/// Trait for all "leaf" nodes: nodes with no ast children.
pub trait Leaf: Node {
/// Returns none if this node is "unsourced." This happens if for whatever reason we are
/// unable to parse the node, either because we had a parse error and recovered, or because
/// we accepted incomplete and the token stream was exhausted.
fn range(&self) -> Option<SourceRange>;
fn range_mut(&mut self) -> &mut Option<SourceRange>;
fn has_source(&self) -> bool {
self.range().is_some()
}
fn leaf_as_node(&self) -> &dyn Node;
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
// A token node is a node which contains a token, which must be one of a fixed set.
pub trait Token: Leaf {
/// The token type which was parsed.
fn token_type(&self) -> ParseTokenType;
fn token_type_mut(&mut self) -> &mut ParseTokenType;
fn allowed_tokens(&self) -> &'static [ParseTokenType];
/// \return whether a token type is allowed in this token_t, i.e. is a member of our Toks list.
fn allows_token(&self, token_type: ParseTokenType) -> bool {
self.allowed_tokens().contains(&token_type)
}
}
/// A keyword node is a node which contains a keyword, which must be one of a fixed set.
pub trait Keyword: Leaf {
fn keyword(&self) -> ParseKeyword;
fn keyword_mut(&mut self) -> &mut ParseKeyword;
fn allowed_keywords(&self) -> &'static [ParseKeyword];
fn allows_keyword(&self, kw: ParseKeyword) -> bool {
self.allowed_keywords().contains(&kw)
}
}
// A simple variable-sized array, possibly empty.
pub trait List: Node {
type ContentsNode: Node + Default;
fn contents(&self) -> &[Box<Self::ContentsNode>];
fn contents_mut(&mut self) -> &mut Vec<Box<Self::ContentsNode>>;
/// \return our count.
fn count(&self) -> usize {
self.contents().len()
}
/// \return whether we are empty.
fn is_empty(&self) -> bool {
self.contents().is_empty()
}
/// Iteration support.
fn iter(&self) -> std::slice::Iter<Box<Self::ContentsNode>> {
self.contents().iter()
}
fn get(&self, index: usize) -> Option<&Self::ContentsNode> {
self.contents().get(index).map(|b| &**b)
}
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
/// Implement the node trait.
macro_rules! implement_node {
(
$name:ident,
$category:ident,
$type:ident $(,)?
) => {
impl Node for $name {
fn typ(&self) -> Type {
Type::$type
}
fn parent(&self) -> Option<&dyn Node> {
self.parent.map(|p| unsafe { &*p })
}
fn category(&self) -> Category {
Category::$category
}
fn try_source_range(&self) -> Option<SourceRange> {
let mut visitor = SourceRangeVisitor {
total: SourceRange::new(0, 0),
any_unsourced: false,
};
visitor.visit(self);
if visitor.any_unsourced {
None
} else {
Some(visitor.total)
}
}
fn as_ptr(&self) -> *const () {
(self as *const $name).cast()
}
fn as_node(&self) -> &dyn Node {
self
}
}
impl NodeMut for $name {}
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
};
}
/// Implement the leaf trait.
macro_rules! implement_leaf {
( $name:ident ) => {
impl Leaf for $name {
fn range(&self) -> Option<SourceRange> {
self.range
}
fn range_mut(&mut self) -> &mut Option<SourceRange> {
&mut self.range
}
fn leaf_as_node(&self) -> &dyn Node {
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
self
}
}
impl Acceptor for $name {
#[allow(unused_variables)]
fn accept<'a>(&'a self, visitor: &mut dyn NodeVisitor<'a>, reversed: bool) {}
}
impl AcceptorMut for $name {
#[allow(unused_variables)]
fn accept_mut(&mut self, visitor: &mut dyn NodeVisitorMut, reversed: bool) {
visitor.will_visit_fields_of(self);
visitor.did_visit_fields_of(self, VisitResult::Continue(()));
}
}
impl $name {
/// Set the parent fields of all nodes in the tree rooted at \p self.
fn set_parents(&mut self) {}
}
};
}
/// Define a node that implements the keyword trait.
macro_rules! define_keyword_node {
( $name:ident, $($allowed:expr),* $(,)? ) => {
#[derive(Default, Debug)]
pub struct $name {
parent: Option<*const dyn Node>,
range: Option<SourceRange>,
keyword: ParseKeyword,
}
implement_node!($name, leaf, keyword_base);
implement_leaf!($name);
impl ConcreteNode for $name {
fn as_leaf(&self) -> Option<&dyn Leaf> {
Some(self)
}
fn as_keyword(&self) -> Option<&dyn Keyword> {
Some(self)
}
}
impl ConcreteNodeMut for $name {
fn as_mut_leaf(&mut self) -> Option<&mut dyn Leaf> {
Some(self)
}
fn as_mut_keyword(&mut self) -> Option<&mut dyn Keyword> {
Some(self)
}
}
impl Keyword for $name {
fn keyword(&self) -> ParseKeyword {
self.keyword
}
fn keyword_mut(&mut self) -> &mut ParseKeyword {
&mut self.keyword
}
fn allowed_keywords(&self) -> &'static [ParseKeyword] {
&[$($allowed),*]
}
}
}
}
/// Define a node that implements the token trait.
macro_rules! define_token_node {
( $name:ident, $($allowed:expr),* $(,)? ) => {
#[derive(Default, Debug)]
pub struct $name {
parent: Option<*const dyn Node>,
range: Option<SourceRange>,
parse_token_type: ParseTokenType,
}
implement_node!($name, leaf, token_base);
implement_leaf!($name);
impl ConcreteNode for $name {
fn as_leaf(&self) -> Option<&dyn Leaf> {
Some(self)
}
fn as_token(&self) -> Option<&dyn Token> {
Some(self)
}
}
impl ConcreteNodeMut for $name {
fn as_mut_leaf(&mut self) -> Option<&mut dyn Leaf> {
Some(self)
}
fn as_mut_token(&mut self) -> Option<&mut dyn Token> {
Some(self)
}
}
impl Token for $name {
fn token_type(&self) -> ParseTokenType {
self.parse_token_type
}
fn token_type_mut(&mut self) -> &mut ParseTokenType {
&mut self.parse_token_type
}
fn allowed_tokens(&self) -> &'static [ParseTokenType] {
&[$($allowed),*]
}
}
}
}
/// Define a node that implements the list trait.
macro_rules! define_list_node {
(
$name:ident,
$type:tt,
$contents:ident
) => {
#[derive(Default, Debug)]
pub struct $name {
parent: Option<*const dyn Node>,
list_contents: Vec<Box<$contents>>,
}
implement_node!($name, list, $type);
impl List for $name {
type ContentsNode = $contents;
fn contents(&self) -> &[Box<Self::ContentsNode>] {
&self.list_contents
}
fn contents_mut(&mut self) -> &mut Vec<Box<Self::ContentsNode>> {
&mut self.list_contents
}
}
impl<'a> IntoIterator for &'a $name {
type Item = &'a Box<$contents>;
type IntoIter = std::slice::Iter<'a, Box<$contents>>;
fn into_iter(self) -> Self::IntoIter {
self.contents().into_iter()
}
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
impl Index<usize> for $name {
type Output = <$name as List>::ContentsNode;
fn index(&self, index: usize) -> &Self::Output {
&*self.contents()[index]
}
}
impl IndexMut<usize> for $name {
fn index_mut(&mut self, index: usize) -> &mut Self::Output {
&mut *self.contents_mut()[index]
}
}
impl Acceptor for $name {
#[allow(unused_variables)]
fn accept<'a>(&'a self, visitor: &mut dyn NodeVisitor<'a>, reversed: bool) {
accept_list_visitor!(Self, accept, visit, self, visitor, reversed, $contents);
}
}
impl AcceptorMut for $name {
#[allow(unused_variables)]
fn accept_mut(&mut self, visitor: &mut dyn NodeVisitorMut, reversed: bool) {
visitor.will_visit_fields_of(self);
let flow = accept_list_visitor!(
Self, accept_mut, visit_mut, self, visitor, reversed, $contents
);
visitor.did_visit_fields_of(self, flow);
}
}
impl $name {
/// Set the parent fields of all nodes in the tree rooted at \p self.
fn set_parents(&mut self) {
for i in 0..self.count() {
self[i].parent = Some(self);
self[i].set_parents();
}
}
}
};
}
macro_rules! accept_list_visitor {
(
$Self:ident,
$accept:ident,
$visit:ident,
$self:ident,
$visitor:ident,
$reversed:ident,
$list_element:ident
) => {
loop {
let mut result = VisitResult::Continue(());
// list types pretend their child nodes are direct embeddings.
// This isn't used during AST construction because we need to construct the list.
if $reversed {
for i in (0..$self.count()).rev() {
result = accept_list_visitor_impl!($self, $visitor, $visit, $self[i]);
if result.is_break() {
break;
}
}
} else {
for i in 0..$self.count() {
result = accept_list_visitor_impl!($self, $visitor, $visit, $self[i]);
if result.is_break() {
break;
}
}
}
break result;
}
};
}
macro_rules! accept_list_visitor_impl {
(
$self:ident,
$visitor:ident,
visit,
$child:expr) => {{
$visitor.visit(&$child);
VisitResult::Continue(())
}};
(
$self:ident,
$visitor:ident,
visit_mut,
$child:expr) => {
$visitor.visit_mut(&mut $child)
};
}
/// Implement the acceptor trait for the given branch node.
macro_rules! implement_acceptor_for_branch {
(
$name:ident
$(, ($field_name:ident: $field_type:tt) )*
$(,)?
) => {
impl Acceptor for $name {
#[allow(unused_variables)]
fn accept<'a>(&'a self, visitor: &mut dyn NodeVisitor<'a>, reversed: bool){
visitor_accept_field!(
Self,
accept,
visit,
self,
visitor,
reversed,
( $( $field_name: $field_type, )* ) );
}
}
impl AcceptorMut for $name {
#[allow(unused_variables)]
fn accept_mut(&mut self, visitor: &mut dyn NodeVisitorMut, reversed: bool) {
visitor.will_visit_fields_of(self);
let flow = visitor_accept_field!(
Self,
accept_mut,
visit_mut,
self,
visitor,
reversed,
( $( $field_name: $field_type, )* ));
visitor.did_visit_fields_of(self, flow);
}
}
impl $name {
/// Set the parent fields of all nodes in the tree rooted at \p self.
fn set_parents(&mut self) {
$(
set_parent_of_field!(self, $field_name, $field_type);
)*
}
}
}
}
2023-11-19 00:51:52 +08:00
/// Visit the given fields in order, returning whether the visitation succeeded.
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
macro_rules! visitor_accept_field {
(
$Self:ident,
$accept:ident,
$visit:ident,
$self:ident,
$visitor:ident,
$reversed:ident,
$fields:tt
) => {
loop {
visitor_accept_field_impl!($visit, $self, $visitor, $reversed, $fields);
break VisitResult::Continue(());
}
};
}
/// Visit the given fields in order, breaking if a visitation fails.
macro_rules! visitor_accept_field_impl {
// Base case: no fields left to visit.
(
$visit:ident,
$self:ident,
$visitor:ident,
$reversed:ident,
()
) => {};
// Visit the first or last field and then the rest.
(
$visit:ident,
$self:ident,
$visitor:ident,
$reversed:ident,
(
$field_name:ident: $field_type:tt,
$( $field_names:ident: $field_types:tt, )*
)
) => {
if !$reversed {
visit_1_field!($visit, ($self.$field_name), $field_type, $visitor);
}
visitor_accept_field_impl!(
$visit, $self, $visitor, $reversed,
( $( $field_names: $field_types, )* ));
if $reversed {
visit_1_field!($visit, ($self.$field_name), $field_type, $visitor);
}
}
}
/// Visit the given field, breaking on failure.
macro_rules! visit_1_field {
(
visit,
$field:expr,
$field_type:tt,
$visitor:ident
) => {
visit_1_field_impl!(visit, $field, $field_type, $visitor);
};
(
visit_mut,
$field:expr,
$field_type:tt,
$visitor:ident
) => {
let result = visit_1_field_impl!(visit_mut, $field, $field_type, $visitor);
if result.is_break() {
break result;
}
};
}
/// Visit the given field.
macro_rules! visit_1_field_impl {
(
$visit:ident,
$field:expr,
(Box<$field_type:ident>),
$visitor:ident
) => {
visit_union_field!($visit, $field_type, $field, $visitor)
};
(
$visit:ident,
$field:expr,
(Option<$field_type:ident>),
$visitor:ident
) => {
visit_optional_field!($visit, $field_type, $field, $visitor)
};
(
$visit:ident,
$field:expr,
$field_type:tt,
$visitor:ident
) => {
$visitor.$visit(apply_borrow!($visit, $field))
};
}
macro_rules! apply_borrow {
( visit, $expr:expr ) => {
&$expr
};
( visit_mut, $expr:expr ) => {
&mut $expr
};
}
macro_rules! visit_union_field {
(
visit,
$field_type:ident,
$field:expr,
$visitor:ident
) => {
$visitor.visit($field.embedded_node().as_node())
};
(
visit_mut,
$field_type:ident,
$field:expr,
$visitor:ident
) => {
visit_union_field_mut!($field_type, $visitor, $field)
};
}
macro_rules! visit_union_field_mut {
(ArgumentOrRedirectionVariant, $visitor:ident, $field:expr) => {
$visitor.visit_argument_or_redirection(&mut $field)
};
(BlockStatementHeaderVariant, $visitor:ident, $field:expr) => {
$visitor.visit_block_statement_header(&mut $field)
};
(StatementVariant, $visitor:ident, $field:expr) => {
$visitor.visit_statement(&mut $field)
};
}
macro_rules! visit_optional_field {
(
visit,
$field_type:ident,
$field:expr,
$visitor:ident
) => {
match &$field {
Some(value) => $visitor.visit(&*value),
None => visit_result!(visit),
}
};
(
visit_mut,
$field_type:ident,
$field:expr,
$visitor:ident
) => {{
visit_optional_field_mut!($field_type, $field, $visitor);
VisitResult::Continue(())
}};
}
macro_rules! visit_optional_field_mut {
(DecoratedStatementDecorator, $field:expr, $visitor:ident) => {
$visitor.visit_decorated_statement_decorator(&mut $field);
};
(JobConjunctionDecorator, $field:expr, $visitor:ident) => {
$visitor.visit_job_conjunction_decorator(&mut $field);
};
(ElseClause, $field:expr, $visitor:ident) => {
$visitor.visit_else_clause(&mut $field);
};
(SemiNl, $field:expr, $visitor:ident) => {
$visitor.visit_semi_nl(&mut $field);
};
(KeywordTime, $field:expr, $visitor:ident) => {
$visitor.visit_time(&mut $field);
};
(TokenBackground, $field:expr, $visitor:ident) => {
$visitor.visit_token_background(&mut $field);
};
}
macro_rules! visit_result {
( visit) => {
()
};
( visit_mut ) => {
VisitResult::Continue(())
};
}
macro_rules! set_parent_of_field {
(
$self:ident,
$field_name:ident,
(Box<$field_type:ident>)
) => {
set_parent_of_union_field!($self, $field_name, $field_type);
};
(
$self:ident,
$field_name:ident,
(Option<$field_type:ident>)
) => {
if $self.$field_name.is_some() {
$self.$field_name.as_mut().unwrap().parent = Some($self);
$self.$field_name.as_mut().unwrap().set_parents();
}
};
(
$self:ident,
$field_name:ident,
$field_type:tt
) => {
$self.$field_name.parent = Some($self);
$self.$field_name.set_parents();
};
}
macro_rules! set_parent_of_union_field {
(
$self:ident,
$field_name:ident,
ArgumentOrRedirectionVariant
) => {
if matches!(
*$self.$field_name,
ArgumentOrRedirectionVariant::Argument(_)
) {
$self.$field_name.as_mut_argument().parent = Some($self);
$self.$field_name.as_mut_argument().set_parents();
} else {
$self.$field_name.as_mut_redirection().parent = Some($self);
$self.$field_name.as_mut_redirection().set_parents();
}
};
(
$self:ident,
$field_name:ident,
StatementVariant
) => {
if matches!(*$self.$field_name, StatementVariant::NotStatement(_)) {
$self.$field_name.as_mut_not_statement().parent = Some($self);
$self.$field_name.as_mut_not_statement().set_parents();
} else if matches!(*$self.$field_name, StatementVariant::BlockStatement(_)) {
$self.$field_name.as_mut_block_statement().parent = Some($self);
$self.$field_name.as_mut_block_statement().set_parents();
} else if matches!(*$self.$field_name, StatementVariant::IfStatement(_)) {
$self.$field_name.as_mut_if_statement().parent = Some($self);
$self.$field_name.as_mut_if_statement().set_parents();
} else if matches!(*$self.$field_name, StatementVariant::SwitchStatement(_)) {
$self.$field_name.as_mut_switch_statement().parent = Some($self);
$self.$field_name.as_mut_switch_statement().set_parents();
} else if matches!(*$self.$field_name, StatementVariant::DecoratedStatement(_)) {
$self.$field_name.as_mut_decorated_statement().parent = Some($self);
$self.$field_name.as_mut_decorated_statement().set_parents();
}
};
(
$self:ident,
$field_name:ident,
BlockStatementHeaderVariant
) => {
if matches!(
*$self.$field_name,
BlockStatementHeaderVariant::ForHeader(_)
) {
$self.$field_name.as_mut_for_header().parent = Some($self);
$self.$field_name.as_mut_for_header().set_parents();
} else if matches!(
*$self.$field_name,
BlockStatementHeaderVariant::WhileHeader(_)
) {
$self.$field_name.as_mut_while_header().parent = Some($self);
$self.$field_name.as_mut_while_header().set_parents();
} else if matches!(
*$self.$field_name,
BlockStatementHeaderVariant::FunctionHeader(_)
) {
$self.$field_name.as_mut_function_header().parent = Some($self);
$self.$field_name.as_mut_function_header().set_parents();
} else if matches!(
*$self.$field_name,
BlockStatementHeaderVariant::BeginHeader(_)
) {
$self.$field_name.as_mut_begin_header().parent = Some($self);
$self.$field_name.as_mut_begin_header().set_parents();
}
};
}
/// A redirection has an operator like > or 2>, and a target like /dev/null or &1.
/// Note that pipes are not redirections.
#[derive(Default, Debug)]
pub struct Redirection {
parent: Option<*const dyn Node>,
pub oper: TokenRedirection,
pub target: String_,
}
implement_node!(Redirection, branch, redirection);
implement_acceptor_for_branch!(Redirection, (oper: TokenRedirection), (target: String_));
impl ConcreteNode for Redirection {
fn as_redirection(&self) -> Option<&Redirection> {
Some(self)
}
}
impl ConcreteNodeMut for Redirection {
fn as_mut_redirection(&mut self) -> Option<&mut Redirection> {
Some(self)
}
}
define_list_node!(
VariableAssignmentList,
variable_assignment_list,
VariableAssignment
);
impl ConcreteNode for VariableAssignmentList {
fn as_variable_assignment_list(&self) -> Option<&VariableAssignmentList> {
Some(self)
}
}
impl ConcreteNodeMut for VariableAssignmentList {
fn as_mut_variable_assignment_list(&mut self) -> Option<&mut VariableAssignmentList> {
Some(self)
}
}
/// An argument or redirection holds either an argument or redirection.
#[derive(Default, Debug)]
pub struct ArgumentOrRedirection {
parent: Option<*const dyn Node>,
pub contents: Box<ArgumentOrRedirectionVariant>,
}
implement_node!(ArgumentOrRedirection, branch, argument_or_redirection);
implement_acceptor_for_branch!(
ArgumentOrRedirection,
(contents: (Box<ArgumentOrRedirectionVariant>))
);
impl ConcreteNode for ArgumentOrRedirection {
fn as_argument_or_redirection(&self) -> Option<&ArgumentOrRedirection> {
Some(self)
}
}
impl ConcreteNodeMut for ArgumentOrRedirection {
fn as_mut_argument_or_redirection(&mut self) -> Option<&mut ArgumentOrRedirection> {
Some(self)
}
}
define_list_node!(
ArgumentOrRedirectionList,
argument_or_redirection_list,
ArgumentOrRedirection
);
impl ConcreteNode for ArgumentOrRedirectionList {
fn as_argument_or_redirection_list(&self) -> Option<&ArgumentOrRedirectionList> {
Some(self)
}
}
impl ConcreteNodeMut for ArgumentOrRedirectionList {
fn as_mut_argument_or_redirection_list(&mut self) -> Option<&mut ArgumentOrRedirectionList> {
Some(self)
}
}
/// A statement is a normal command, or an if / while / etc
#[derive(Default, Debug)]
pub struct Statement {
parent: Option<*const dyn Node>,
pub contents: Box<StatementVariant>,
}
implement_node!(Statement, branch, statement);
implement_acceptor_for_branch!(Statement, (contents: (Box<StatementVariant>)));
impl ConcreteNode for Statement {
fn as_statement(&self) -> Option<&Statement> {
Some(self)
}
}
impl ConcreteNodeMut for Statement {
fn as_mut_statement(&mut self) -> Option<&mut Statement> {
Some(self)
}
}
/// A job is a non-empty list of statements, separated by pipes. (Non-empty is useful for cases
/// like if statements, where we require a command).
#[derive(Default, Debug)]
pub struct JobPipeline {
parent: Option<*const dyn Node>,
/// Maybe the time keyword.
pub time: Option<KeywordTime>,
/// A (possibly empty) list of variable assignments.
pub variables: VariableAssignmentList,
/// The statement.
pub statement: Statement,
/// Piped remainder.
pub continuation: JobContinuationList,
/// Maybe backgrounded.
pub bg: Option<TokenBackground>,
}
implement_node!(JobPipeline, branch, job_pipeline);
implement_acceptor_for_branch!(
JobPipeline,
(time: (Option<KeywordTime>)),
(variables: (VariableAssignmentList)),
(statement: (Statement)),
(continuation: (JobContinuationList)),
(bg: (Option<TokenBackground>)),
);
impl ConcreteNode for JobPipeline {
fn as_job_pipeline(&self) -> Option<&JobPipeline> {
Some(self)
}
}
impl ConcreteNodeMut for JobPipeline {
fn as_mut_job_pipeline(&mut self) -> Option<&mut JobPipeline> {
Some(self)
}
}
/// A job_conjunction is a job followed by a && or || continuations.
#[derive(Default, Debug)]
pub struct JobConjunction {
parent: Option<*const dyn Node>,
/// The job conjunction decorator.
pub decorator: Option<JobConjunctionDecorator>,
/// The job itself.
pub job: JobPipeline,
/// The rest of the job conjunction, with && or ||s.
pub continuations: JobConjunctionContinuationList,
/// A terminating semicolon or newline. This is marked optional because it may not be
/// present, for example the command `echo foo` may not have a terminating newline. It will
/// only fail to be present if we ran out of tokens.
pub semi_nl: Option<SemiNl>,
}
implement_node!(JobConjunction, branch, job_conjunction);
implement_acceptor_for_branch!(
JobConjunction,
(decorator: (Option<JobConjunctionDecorator>)),
(job: (JobPipeline)),
(continuations: (JobConjunctionContinuationList)),
(semi_nl: (Option<SemiNl>)),
);
impl ConcreteNode for JobConjunction {
fn as_job_conjunction(&self) -> Option<&JobConjunction> {
Some(self)
}
}
impl ConcreteNodeMut for JobConjunction {
fn as_mut_job_conjunction(&mut self) -> Option<&mut JobConjunction> {
Some(self)
}
}
#[derive(Default, Debug)]
pub struct ForHeader {
parent: Option<*const dyn Node>,
/// 'for'
pub kw_for: KeywordFor,
/// var_name
pub var_name: String_,
/// 'in'
pub kw_in: KeywordIn,
/// list of arguments
pub args: ArgumentList,
/// newline or semicolon
pub semi_nl: SemiNl,
}
implement_node!(ForHeader, branch, for_header);
implement_acceptor_for_branch!(
ForHeader,
(kw_for: (KeywordFor)),
(var_name: (String_)),
(kw_in: (KeywordIn)),
(args: (ArgumentList)),
(semi_nl: (SemiNl)),
);
impl ConcreteNode for ForHeader {
fn as_for_header(&self) -> Option<&ForHeader> {
Some(self)
}
}
impl ConcreteNodeMut for ForHeader {
fn as_mut_for_header(&mut self) -> Option<&mut ForHeader> {
Some(self)
}
}
#[derive(Default, Debug)]
pub struct WhileHeader {
parent: Option<*const dyn Node>,
/// 'while'
pub kw_while: KeywordWhile,
pub condition: JobConjunction,
pub andor_tail: AndorJobList,
}
implement_node!(WhileHeader, branch, while_header);
implement_acceptor_for_branch!(
WhileHeader,
(kw_while: (KeywordWhile)),
(condition: (JobConjunction)),
(andor_tail: (AndorJobList)),
);
impl ConcreteNode for WhileHeader {
fn as_while_header(&self) -> Option<&WhileHeader> {
Some(self)
}
}
impl ConcreteNodeMut for WhileHeader {
fn as_mut_while_header(&mut self) -> Option<&mut WhileHeader> {
Some(self)
}
}
#[derive(Default, Debug)]
pub struct FunctionHeader {
parent: Option<*const dyn Node>,
pub kw_function: KeywordFunction,
/// functions require at least one argument.
pub first_arg: Argument,
pub args: ArgumentList,
pub semi_nl: SemiNl,
}
implement_node!(FunctionHeader, branch, function_header);
implement_acceptor_for_branch!(
FunctionHeader,
(kw_function: (KeywordFunction)),
(first_arg: (Argument)),
(args: (ArgumentList)),
(semi_nl: (SemiNl)),
);
impl ConcreteNode for FunctionHeader {
fn as_function_header(&self) -> Option<&FunctionHeader> {
Some(self)
}
}
impl ConcreteNodeMut for FunctionHeader {
fn as_mut_function_header(&mut self) -> Option<&mut FunctionHeader> {
Some(self)
}
}
#[derive(Default, Debug)]
pub struct BeginHeader {
parent: Option<*const dyn Node>,
pub kw_begin: KeywordBegin,
/// Note that 'begin' does NOT require a semi or nl afterwards.
/// This is valid: begin echo hi; end
pub semi_nl: Option<SemiNl>,
}
implement_node!(BeginHeader, branch, begin_header);
implement_acceptor_for_branch!(
BeginHeader,
(kw_begin: (KeywordBegin)),
(semi_nl: (Option<SemiNl>))
);
impl ConcreteNode for BeginHeader {
fn as_begin_header(&self) -> Option<&BeginHeader> {
Some(self)
}
}
impl ConcreteNodeMut for BeginHeader {
fn as_mut_begin_header(&mut self) -> Option<&mut BeginHeader> {
Some(self)
}
}
#[derive(Default, Debug)]
pub struct BlockStatement {
parent: Option<*const dyn Node>,
/// A header like for, while, etc.
pub header: Box<BlockStatementHeaderVariant>,
/// List of jobs in this block.
pub jobs: JobList,
/// The 'end' node.
pub end: KeywordEnd,
/// Arguments and redirections associated with the block.
pub args_or_redirs: ArgumentOrRedirectionList,
}
implement_node!(BlockStatement, branch, block_statement);
implement_acceptor_for_branch!(
BlockStatement,
(header: (Box<BlockStatementHeaderVariant>)),
(jobs: (JobList)),
(end: (KeywordEnd)),
(args_or_redirs: (ArgumentOrRedirectionList)),
);
impl ConcreteNode for BlockStatement {
fn as_block_statement(&self) -> Option<&BlockStatement> {
Some(self)
}
}
impl ConcreteNodeMut for BlockStatement {
fn as_mut_block_statement(&mut self) -> Option<&mut BlockStatement> {
Some(self)
}
}
#[derive(Default, Debug)]
pub struct IfClause {
parent: Option<*const dyn Node>,
/// The 'if' keyword.
pub kw_if: KeywordIf,
/// The 'if' condition.
pub condition: JobConjunction,
/// 'and/or' tail.
pub andor_tail: AndorJobList,
/// The body to execute if the condition is true.
pub body: JobList,
}
implement_node!(IfClause, branch, if_clause);
implement_acceptor_for_branch!(
IfClause,
(kw_if: (KeywordIf)),
(condition: (JobConjunction)),
(andor_tail: (AndorJobList)),
(body: (JobList)),
);
impl ConcreteNode for IfClause {
fn as_if_clause(&self) -> Option<&IfClause> {
Some(self)
}
}
impl ConcreteNodeMut for IfClause {
fn as_mut_if_clause(&mut self) -> Option<&mut IfClause> {
Some(self)
}
}
#[derive(Default, Debug)]
pub struct ElseifClause {
parent: Option<*const dyn Node>,
/// The 'else' keyword.
pub kw_else: KeywordElse,
/// The 'if' clause following it.
pub if_clause: IfClause,
}
implement_node!(ElseifClause, branch, elseif_clause);
implement_acceptor_for_branch!(
ElseifClause,
(kw_else: (KeywordElse)),
(if_clause: (IfClause)),
);
impl ConcreteNode for ElseifClause {
fn as_elseif_clause(&self) -> Option<&ElseifClause> {
Some(self)
}
}
impl ConcreteNodeMut for ElseifClause {
fn as_mut_elseif_clause(&mut self) -> Option<&mut ElseifClause> {
Some(self)
}
}
define_list_node!(ElseifClauseList, elseif_clause_list, ElseifClause);
impl ConcreteNode for ElseifClauseList {
fn as_elseif_clause_list(&self) -> Option<&ElseifClauseList> {
Some(self)
}
}
impl ConcreteNodeMut for ElseifClauseList {
fn as_mut_elseif_clause_list(&mut self) -> Option<&mut ElseifClauseList> {
Some(self)
}
}
#[derive(Default, Debug)]
pub struct ElseClause {
parent: Option<*const dyn Node>,
/// else ; body
pub kw_else: KeywordElse,
pub semi_nl: SemiNl,
pub body: JobList,
}
implement_node!(ElseClause, branch, else_clause);
implement_acceptor_for_branch!(
ElseClause,
(kw_else: (KeywordElse)),
(semi_nl: (SemiNl)),
(body: (JobList)),
);
impl ConcreteNode for ElseClause {
fn as_else_clause(&self) -> Option<&ElseClause> {
Some(self)
}
}
impl ConcreteNodeMut for ElseClause {
fn as_mut_else_clause(&mut self) -> Option<&mut ElseClause> {
Some(self)
}
}
#[derive(Default, Debug)]
pub struct IfStatement {
parent: Option<*const dyn Node>,
/// if part
pub if_clause: IfClause,
/// else if list
pub elseif_clauses: ElseifClauseList,
/// else part
pub else_clause: Option<ElseClause>,
/// literal end
pub end: KeywordEnd,
/// block args / redirs
pub args_or_redirs: ArgumentOrRedirectionList,
}
implement_node!(IfStatement, branch, if_statement);
implement_acceptor_for_branch!(
IfStatement,
(if_clause: (IfClause)),
(elseif_clauses: (ElseifClauseList)),
(else_clause: (Option<ElseClause>)),
(end: (KeywordEnd)),
(args_or_redirs: (ArgumentOrRedirectionList)),
);
impl ConcreteNode for IfStatement {
fn as_if_statement(&self) -> Option<&IfStatement> {
Some(self)
}
}
impl ConcreteNodeMut for IfStatement {
fn as_mut_if_statement(&mut self) -> Option<&mut IfStatement> {
Some(self)
}
}
#[derive(Default, Debug)]
pub struct CaseItem {
parent: Option<*const dyn Node>,
/// case <arguments> ; body
pub kw_case: KeywordCase,
pub arguments: ArgumentList,
pub semi_nl: SemiNl,
pub body: JobList,
}
implement_node!(CaseItem, branch, case_item);
implement_acceptor_for_branch!(
CaseItem,
(kw_case: (KeywordCase)),
(arguments: (ArgumentList)),
(semi_nl: (SemiNl)),
(body: (JobList)),
);
impl ConcreteNode for CaseItem {
fn as_case_item(&self) -> Option<&CaseItem> {
Some(self)
}
}
impl ConcreteNodeMut for CaseItem {
fn as_mut_case_item(&mut self) -> Option<&mut CaseItem> {
Some(self)
}
}
#[derive(Default, Debug)]
pub struct SwitchStatement {
parent: Option<*const dyn Node>,
/// switch <argument> ; body ; end args_redirs
pub kw_switch: KeywordSwitch,
pub argument: Argument,
pub semi_nl: SemiNl,
pub cases: CaseItemList,
pub end: KeywordEnd,
pub args_or_redirs: ArgumentOrRedirectionList,
}
implement_node!(SwitchStatement, branch, switch_statement);
implement_acceptor_for_branch!(
SwitchStatement,
(kw_switch: (KeywordSwitch)),
(argument: (Argument)),
(semi_nl: (SemiNl)),
(cases: (CaseItemList)),
(end: (KeywordEnd)),
(args_or_redirs: (ArgumentOrRedirectionList)),
);
impl ConcreteNode for SwitchStatement {
fn as_switch_statement(&self) -> Option<&SwitchStatement> {
Some(self)
}
}
impl ConcreteNodeMut for SwitchStatement {
fn as_mut_switch_statement(&mut self) -> Option<&mut SwitchStatement> {
Some(self)
}
}
/// A decorated_statement is a command with a list of arguments_or_redirections, possibly with
/// "builtin" or "command" or "exec"
#[derive(Default, Debug)]
pub struct DecoratedStatement {
parent: Option<*const dyn Node>,
/// An optional decoration (command, builtin, exec, etc).
pub opt_decoration: Option<DecoratedStatementDecorator>,
/// Command to run.
pub command: String_,
/// Args and redirs
pub args_or_redirs: ArgumentOrRedirectionList,
}
implement_node!(DecoratedStatement, branch, decorated_statement);
implement_acceptor_for_branch!(
DecoratedStatement,
(opt_decoration: (Option<DecoratedStatementDecorator>)),
(command: (String_)),
(args_or_redirs: (ArgumentOrRedirectionList)),
);
impl ConcreteNode for DecoratedStatement {
fn as_decorated_statement(&self) -> Option<&DecoratedStatement> {
Some(self)
}
}
impl ConcreteNodeMut for DecoratedStatement {
fn as_mut_decorated_statement(&mut self) -> Option<&mut DecoratedStatement> {
Some(self)
}
}
/// A not statement like `not true` or `! true`
#[derive(Default, Debug)]
pub struct NotStatement {
parent: Option<*const dyn Node>,
/// Keyword, either not or exclam.
pub kw: KeywordNot,
pub variables: VariableAssignmentList,
pub time: Option<KeywordTime>,
pub contents: Statement,
}
implement_node!(NotStatement, branch, not_statement);
implement_acceptor_for_branch!(
NotStatement,
(kw: (KeywordNot)),
(variables: (VariableAssignmentList)),
(time: (Option<KeywordTime>)),
(contents: (Statement)),
);
impl ConcreteNode for NotStatement {
fn as_not_statement(&self) -> Option<&NotStatement> {
Some(self)
}
}
impl ConcreteNodeMut for NotStatement {
fn as_mut_not_statement(&mut self) -> Option<&mut NotStatement> {
Some(self)
}
}
#[derive(Default, Debug)]
pub struct JobContinuation {
parent: Option<*const dyn Node>,
pub pipe: TokenPipe,
pub newlines: MaybeNewlines,
pub variables: VariableAssignmentList,
pub statement: Statement,
}
implement_node!(JobContinuation, branch, job_continuation);
implement_acceptor_for_branch!(
JobContinuation,
(pipe: (TokenPipe)),
(newlines: (MaybeNewlines)),
(variables: (VariableAssignmentList)),
(statement: (Statement)),
);
impl ConcreteNode for JobContinuation {
fn as_job_continuation(&self) -> Option<&JobContinuation> {
Some(self)
}
}
impl ConcreteNodeMut for JobContinuation {
fn as_mut_job_continuation(&mut self) -> Option<&mut JobContinuation> {
Some(self)
}
}
define_list_node!(JobContinuationList, job_continuation_list, JobContinuation);
impl ConcreteNode for JobContinuationList {
fn as_job_continuation_list(&self) -> Option<&JobContinuationList> {
Some(self)
}
}
impl ConcreteNodeMut for JobContinuationList {
fn as_mut_job_continuation_list(&mut self) -> Option<&mut JobContinuationList> {
Some(self)
}
}
#[derive(Default, Debug)]
pub struct JobConjunctionContinuation {
parent: Option<*const dyn Node>,
/// The && or || token.
pub conjunction: TokenConjunction,
pub newlines: MaybeNewlines,
/// The job itself.
pub job: JobPipeline,
}
implement_node!(
JobConjunctionContinuation,
branch,
job_conjunction_continuation
);
implement_acceptor_for_branch!(
JobConjunctionContinuation,
(conjunction: (TokenConjunction)),
(newlines: (MaybeNewlines)),
(job: (JobPipeline)),
);
impl ConcreteNode for JobConjunctionContinuation {
fn as_job_conjunction_continuation(&self) -> Option<&JobConjunctionContinuation> {
Some(self)
}
}
impl ConcreteNodeMut for JobConjunctionContinuation {
fn as_mut_job_conjunction_continuation(&mut self) -> Option<&mut JobConjunctionContinuation> {
Some(self)
}
}
/// An andor_job just wraps a job, but requires that the job have an 'and' or 'or' job_decorator.
/// Note this is only used for andor_job_list; jobs that are not part of an andor_job_list are not
/// instances of this.
#[derive(Default, Debug)]
pub struct AndorJob {
parent: Option<*const dyn Node>,
pub job: JobConjunction,
}
implement_node!(AndorJob, branch, andor_job);
implement_acceptor_for_branch!(AndorJob, (job: (JobConjunction)));
impl ConcreteNode for AndorJob {
fn as_andor_job(&self) -> Option<&AndorJob> {
Some(self)
}
}
impl ConcreteNodeMut for AndorJob {
fn as_mut_andor_job(&mut self) -> Option<&mut AndorJob> {
Some(self)
}
}
define_list_node!(AndorJobList, andor_job_list, AndorJob);
impl ConcreteNode for AndorJobList {
fn as_andor_job_list(&self) -> Option<&AndorJobList> {
Some(self)
}
}
impl ConcreteNodeMut for AndorJobList {
fn as_mut_andor_job_list(&mut self) -> Option<&mut AndorJobList> {
Some(self)
}
}
/// A freestanding_argument_list is equivalent to a normal argument list, except it may contain
/// TOK_END (newlines, and even semicolons, for historical reasons).
/// In practice the tok_ends are ignored by fish code so we do not bother to store them.
#[derive(Default, Debug)]
pub struct FreestandingArgumentList {
parent: Option<*const dyn Node>,
pub arguments: ArgumentList,
}
implement_node!(FreestandingArgumentList, branch, freestanding_argument_list);
implement_acceptor_for_branch!(FreestandingArgumentList, (arguments: (ArgumentList)));
impl ConcreteNode for FreestandingArgumentList {
fn as_freestanding_argument_list(&self) -> Option<&FreestandingArgumentList> {
Some(self)
}
}
impl ConcreteNodeMut for FreestandingArgumentList {
fn as_mut_freestanding_argument_list(&mut self) -> Option<&mut FreestandingArgumentList> {
Some(self)
}
}
define_list_node!(
JobConjunctionContinuationList,
job_conjunction_continuation_list,
JobConjunctionContinuation
);
impl ConcreteNode for JobConjunctionContinuationList {
fn as_job_conjunction_continuation_list(&self) -> Option<&JobConjunctionContinuationList> {
Some(self)
}
}
impl ConcreteNodeMut for JobConjunctionContinuationList {
fn as_mut_job_conjunction_continuation_list(
&mut self,
) -> Option<&mut JobConjunctionContinuationList> {
Some(self)
}
}
define_list_node!(ArgumentList, argument_list, Argument);
impl ConcreteNode for ArgumentList {
fn as_argument_list(&self) -> Option<&ArgumentList> {
Some(self)
}
}
impl ConcreteNodeMut for ArgumentList {
fn as_mut_argument_list(&mut self) -> Option<&mut ArgumentList> {
Some(self)
}
}
// For historical reasons, a job list is a list of job *conjunctions*. This should be fixed.
define_list_node!(JobList, job_list, JobConjunction);
impl ConcreteNode for JobList {
fn as_job_list(&self) -> Option<&JobList> {
Some(self)
}
}
impl ConcreteNodeMut for JobList {
fn as_mut_job_list(&mut self) -> Option<&mut JobList> {
Some(self)
}
}
define_list_node!(CaseItemList, case_item_list, CaseItem);
impl ConcreteNode for CaseItemList {
fn as_case_item_list(&self) -> Option<&CaseItemList> {
Some(self)
}
}
impl ConcreteNodeMut for CaseItemList {
fn as_mut_case_item_list(&mut self) -> Option<&mut CaseItemList> {
Some(self)
}
}
/// A variable_assignment contains a source range like FOO=bar.
#[derive(Default, Debug)]
pub struct VariableAssignment {
parent: Option<*const dyn Node>,
range: Option<SourceRange>,
}
implement_node!(VariableAssignment, leaf, variable_assignment);
implement_leaf!(VariableAssignment);
impl ConcreteNode for VariableAssignment {
fn as_leaf(&self) -> Option<&dyn Leaf> {
Some(self)
}
fn as_variable_assignment(&self) -> Option<&VariableAssignment> {
Some(self)
}
}
impl ConcreteNodeMut for VariableAssignment {
fn as_mut_variable_assignment(&mut self) -> Option<&mut VariableAssignment> {
Some(self)
}
}
/// Zero or more newlines.
#[derive(Default, Debug)]
pub struct MaybeNewlines {
parent: Option<*const dyn Node>,
range: Option<SourceRange>,
}
implement_node!(MaybeNewlines, leaf, maybe_newlines);
implement_leaf!(MaybeNewlines);
impl ConcreteNode for MaybeNewlines {
fn as_leaf(&self) -> Option<&dyn Leaf> {
Some(self)
}
fn as_maybe_newlines(&self) -> Option<&MaybeNewlines> {
Some(self)
}
}
impl ConcreteNodeMut for MaybeNewlines {
fn as_mut_leaf(&mut self) -> Option<&mut dyn Leaf> {
Some(self)
}
fn as_mut_maybe_newlines(&mut self) -> Option<&mut MaybeNewlines> {
Some(self)
}
}
/// An argument is just a node whose source range determines its contents.
/// This is a separate type because it is sometimes useful to find all arguments.
#[derive(Default, Debug)]
pub struct Argument {
parent: Option<*const dyn Node>,
range: Option<SourceRange>,
}
implement_node!(Argument, leaf, argument);
implement_leaf!(Argument);
impl ConcreteNode for Argument {
fn as_leaf(&self) -> Option<&dyn Leaf> {
Some(self)
}
fn as_argument(&self) -> Option<&Argument> {
Some(self)
}
}
impl ConcreteNodeMut for Argument {
fn as_mut_leaf(&mut self) -> Option<&mut dyn Leaf> {
Some(self)
}
fn as_mut_argument(&mut self) -> Option<&mut Argument> {
Some(self)
}
}
define_token_node!(SemiNl, ParseTokenType::end);
define_token_node!(String_, ParseTokenType::string);
define_token_node!(TokenBackground, ParseTokenType::background);
#[rustfmt::skip]
define_token_node!(TokenConjunction, ParseTokenType::andand, ParseTokenType::oror);
define_token_node!(TokenPipe, ParseTokenType::pipe);
define_token_node!(TokenRedirection, ParseTokenType::redirection);
#[rustfmt::skip]
define_keyword_node!(DecoratedStatementDecorator, ParseKeyword::kw_command, ParseKeyword::kw_builtin, ParseKeyword::kw_exec);
#[rustfmt::skip]
define_keyword_node!(JobConjunctionDecorator, ParseKeyword::kw_and, ParseKeyword::kw_or);
#[rustfmt::skip]
define_keyword_node!(KeywordBegin, ParseKeyword::kw_begin);
define_keyword_node!(KeywordCase, ParseKeyword::kw_case);
define_keyword_node!(KeywordElse, ParseKeyword::kw_else);
define_keyword_node!(KeywordEnd, ParseKeyword::kw_end);
define_keyword_node!(KeywordFor, ParseKeyword::kw_for);
define_keyword_node!(KeywordFunction, ParseKeyword::kw_function);
define_keyword_node!(KeywordIf, ParseKeyword::kw_if);
define_keyword_node!(KeywordIn, ParseKeyword::kw_in);
#[rustfmt::skip]
define_keyword_node!(KeywordNot, ParseKeyword::kw_not, ParseKeyword::kw_builtin, ParseKeyword::kw_exclam);
define_keyword_node!(KeywordSwitch, ParseKeyword::kw_switch);
define_keyword_node!(KeywordTime, ParseKeyword::kw_time);
define_keyword_node!(KeywordWhile, ParseKeyword::kw_while);
impl DecoratedStatement {
/// \return the decoration for this statement.
2023-04-19 06:11:49 +08:00
pub fn decoration(&self) -> StatementDecoration {
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
let Some(decorator) = &self.opt_decoration else {
return StatementDecoration::none;
};
let decorator: &dyn Keyword = decorator;
match decorator.keyword() {
ParseKeyword::kw_command => StatementDecoration::command,
ParseKeyword::kw_builtin => StatementDecoration::builtin,
ParseKeyword::kw_exec => StatementDecoration::exec,
_ => panic!("Unexpected keyword in statement decoration"),
}
}
}
#[derive(Debug)]
pub enum ArgumentOrRedirectionVariant {
Argument(Argument),
Redirection(Redirection),
}
impl Default for ArgumentOrRedirectionVariant {
fn default() -> Self {
ArgumentOrRedirectionVariant::Argument(Argument::default())
}
}
impl Acceptor for ArgumentOrRedirectionVariant {
fn accept<'a>(&'a self, visitor: &mut dyn NodeVisitor<'a>, reversed: bool) {
match self {
ArgumentOrRedirectionVariant::Argument(child) => child.accept(visitor, reversed),
ArgumentOrRedirectionVariant::Redirection(child) => child.accept(visitor, reversed),
}
}
}
impl AcceptorMut for ArgumentOrRedirectionVariant {
fn accept_mut(&mut self, visitor: &mut dyn NodeVisitorMut, reversed: bool) {
match self {
ArgumentOrRedirectionVariant::Argument(child) => child.accept_mut(visitor, reversed),
ArgumentOrRedirectionVariant::Redirection(child) => child.accept_mut(visitor, reversed),
}
}
}
impl ArgumentOrRedirectionVariant {
2023-04-19 06:11:49 +08:00
pub fn typ(&self) -> Type {
self.embedded_node().typ()
}
pub fn try_source_range(&self) -> Option<SourceRange> {
self.embedded_node().try_source_range()
}
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
fn embedded_node(&self) -> &dyn NodeMut {
match self {
ArgumentOrRedirectionVariant::Argument(node) => node,
ArgumentOrRedirectionVariant::Redirection(node) => node,
}
}
fn as_mut_argument(&mut self) -> &mut Argument {
match self {
ArgumentOrRedirectionVariant::Argument(node) => node,
_ => panic!(),
}
}
fn as_mut_redirection(&mut self) -> &mut Redirection {
match self {
ArgumentOrRedirectionVariant::Redirection(redirection) => redirection,
_ => panic!(),
}
}
}
impl ArgumentOrRedirection {
/// \return whether this represents an argument.
pub fn is_argument(&self) -> bool {
matches!(*self.contents, ArgumentOrRedirectionVariant::Argument(_))
}
/// \return whether this represents a redirection
pub fn is_redirection(&self) -> bool {
matches!(*self.contents, ArgumentOrRedirectionVariant::Redirection(_))
}
/// \return this as an argument, assuming it wraps one.
pub fn argument(&self) -> &Argument {
match *self.contents {
ArgumentOrRedirectionVariant::Argument(ref arg) => arg,
_ => panic!("Is not an argument"),
}
}
/// \return this as an argument, assuming it wraps one.
pub fn redirection(&self) -> &Redirection {
match *self.contents {
ArgumentOrRedirectionVariant::Redirection(ref arg) => arg,
_ => panic!("Is not a redirection"),
}
}
}
#[derive(Debug)]
pub enum StatementVariant {
None,
NotStatement(NotStatement),
BlockStatement(BlockStatement),
IfStatement(IfStatement),
SwitchStatement(SwitchStatement),
DecoratedStatement(DecoratedStatement),
}
impl Default for StatementVariant {
fn default() -> Self {
StatementVariant::None
}
}
impl Acceptor for StatementVariant {
fn accept<'a>(&'a self, visitor: &mut dyn NodeVisitor<'a>, reversed: bool) {
match self {
StatementVariant::None => panic!("cannot visit null statement"),
StatementVariant::NotStatement(node) => node.accept(visitor, reversed),
StatementVariant::BlockStatement(node) => node.accept(visitor, reversed),
StatementVariant::IfStatement(node) => node.accept(visitor, reversed),
StatementVariant::SwitchStatement(node) => node.accept(visitor, reversed),
StatementVariant::DecoratedStatement(node) => node.accept(visitor, reversed),
}
}
}
impl AcceptorMut for StatementVariant {
fn accept_mut(&mut self, visitor: &mut dyn NodeVisitorMut, reversed: bool) {
match self {
StatementVariant::None => panic!("cannot visit null statement"),
StatementVariant::NotStatement(node) => node.accept_mut(visitor, reversed),
StatementVariant::BlockStatement(node) => node.accept_mut(visitor, reversed),
StatementVariant::IfStatement(node) => node.accept_mut(visitor, reversed),
StatementVariant::SwitchStatement(node) => node.accept_mut(visitor, reversed),
StatementVariant::DecoratedStatement(node) => node.accept_mut(visitor, reversed),
}
}
}
impl StatementVariant {
2023-04-19 06:11:49 +08:00
pub fn typ(&self) -> Type {
self.embedded_node().typ()
}
pub fn try_source_range(&self) -> Option<SourceRange> {
self.embedded_node().try_source_range()
}
pub fn as_not_statement(&self) -> Option<&NotStatement> {
match self {
StatementVariant::NotStatement(node) => Some(node),
_ => None,
}
}
pub fn as_block_statement(&self) -> Option<&BlockStatement> {
match self {
StatementVariant::BlockStatement(node) => Some(node),
_ => None,
}
}
pub fn as_if_statement(&self) -> Option<&IfStatement> {
match self {
StatementVariant::IfStatement(node) => Some(node),
_ => None,
}
}
pub fn as_switch_statement(&self) -> Option<&SwitchStatement> {
match self {
StatementVariant::SwitchStatement(node) => Some(node),
_ => None,
}
}
pub fn as_decorated_statement(&self) -> Option<&DecoratedStatement> {
match self {
StatementVariant::DecoratedStatement(node) => Some(node),
_ => None,
}
}
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
fn embedded_node(&self) -> &dyn NodeMut {
match self {
StatementVariant::None => panic!("cannot visit null statement"),
StatementVariant::NotStatement(node) => node,
StatementVariant::BlockStatement(node) => node,
StatementVariant::IfStatement(node) => node,
StatementVariant::SwitchStatement(node) => node,
StatementVariant::DecoratedStatement(node) => node,
}
}
fn as_mut_not_statement(&mut self) -> &mut NotStatement {
match self {
StatementVariant::NotStatement(node) => node,
_ => panic!(),
}
}
fn as_mut_block_statement(&mut self) -> &mut BlockStatement {
match self {
StatementVariant::BlockStatement(node) => node,
_ => panic!(),
}
}
fn as_mut_if_statement(&mut self) -> &mut IfStatement {
match self {
StatementVariant::IfStatement(node) => node,
_ => panic!(),
}
}
fn as_mut_switch_statement(&mut self) -> &mut SwitchStatement {
match self {
StatementVariant::SwitchStatement(node) => node,
_ => panic!(),
}
}
fn as_mut_decorated_statement(&mut self) -> &mut DecoratedStatement {
match self {
StatementVariant::DecoratedStatement(node) => node,
_ => panic!(),
}
}
}
#[derive(Debug)]
pub enum BlockStatementHeaderVariant {
None,
ForHeader(ForHeader),
WhileHeader(WhileHeader),
FunctionHeader(FunctionHeader),
BeginHeader(BeginHeader),
}
impl Default for BlockStatementHeaderVariant {
fn default() -> Self {
BlockStatementHeaderVariant::None
}
}
impl Acceptor for BlockStatementHeaderVariant {
fn accept<'a>(&'a self, visitor: &mut dyn NodeVisitor<'a>, reversed: bool) {
match self {
BlockStatementHeaderVariant::None => panic!("cannot visit null block header"),
BlockStatementHeaderVariant::ForHeader(node) => node.accept(visitor, reversed),
BlockStatementHeaderVariant::WhileHeader(node) => node.accept(visitor, reversed),
BlockStatementHeaderVariant::FunctionHeader(node) => node.accept(visitor, reversed),
BlockStatementHeaderVariant::BeginHeader(node) => node.accept(visitor, reversed),
}
}
}
impl AcceptorMut for BlockStatementHeaderVariant {
fn accept_mut(&mut self, visitor: &mut dyn NodeVisitorMut, reversed: bool) {
match self {
BlockStatementHeaderVariant::None => panic!("cannot visit null block header"),
BlockStatementHeaderVariant::ForHeader(node) => node.accept_mut(visitor, reversed),
BlockStatementHeaderVariant::WhileHeader(node) => node.accept_mut(visitor, reversed),
BlockStatementHeaderVariant::FunctionHeader(node) => node.accept_mut(visitor, reversed),
BlockStatementHeaderVariant::BeginHeader(node) => node.accept_mut(visitor, reversed),
}
}
}
impl BlockStatementHeaderVariant {
2023-04-19 06:11:49 +08:00
pub fn typ(&self) -> Type {
self.embedded_node().typ()
}
pub fn try_source_range(&self) -> Option<SourceRange> {
self.embedded_node().try_source_range()
}
pub fn as_for_header(&self) -> Option<&ForHeader> {
match self {
BlockStatementHeaderVariant::ForHeader(node) => Some(node),
_ => None,
}
}
pub fn as_while_header(&self) -> Option<&WhileHeader> {
match self {
BlockStatementHeaderVariant::WhileHeader(node) => Some(node),
_ => None,
}
}
pub fn as_function_header(&self) -> Option<&FunctionHeader> {
match self {
BlockStatementHeaderVariant::FunctionHeader(node) => Some(node),
_ => None,
}
}
pub fn as_begin_header(&self) -> Option<&BeginHeader> {
match self {
BlockStatementHeaderVariant::BeginHeader(node) => Some(node),
_ => None,
}
}
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
fn embedded_node(&self) -> &dyn NodeMut {
match self {
BlockStatementHeaderVariant::None => panic!("cannot visit null block header"),
BlockStatementHeaderVariant::ForHeader(node) => node,
BlockStatementHeaderVariant::WhileHeader(node) => node,
BlockStatementHeaderVariant::FunctionHeader(node) => node,
BlockStatementHeaderVariant::BeginHeader(node) => node,
}
}
fn as_mut_for_header(&mut self) -> &mut ForHeader {
match self {
BlockStatementHeaderVariant::ForHeader(node) => node,
_ => panic!(),
}
}
fn as_mut_while_header(&mut self) -> &mut WhileHeader {
match self {
BlockStatementHeaderVariant::WhileHeader(node) => node,
_ => panic!(),
}
}
fn as_mut_function_header(&mut self) -> &mut FunctionHeader {
match self {
BlockStatementHeaderVariant::FunctionHeader(node) => node,
_ => panic!(),
}
}
fn as_mut_begin_header(&mut self) -> &mut BeginHeader {
match self {
BlockStatementHeaderVariant::BeginHeader(node) => node,
_ => panic!(),
}
}
}
/// \return a string literal name for an ast type.
pub fn ast_type_to_string(t: Type) -> &'static wstr {
match t {
Type::token_base => L!("token_base"),
Type::keyword_base => L!("keyword_base"),
Type::redirection => L!("redirection"),
Type::variable_assignment => L!("variable_assignment"),
Type::variable_assignment_list => L!("variable_assignment_list"),
Type::argument_or_redirection => L!("argument_or_redirection"),
Type::argument_or_redirection_list => L!("argument_or_redirection_list"),
Type::statement => L!("statement"),
Type::job_pipeline => L!("job_pipeline"),
Type::job_conjunction => L!("job_conjunction"),
Type::for_header => L!("for_header"),
Type::while_header => L!("while_header"),
Type::function_header => L!("function_header"),
Type::begin_header => L!("begin_header"),
Type::block_statement => L!("block_statement"),
Type::if_clause => L!("if_clause"),
Type::elseif_clause => L!("elseif_clause"),
Type::elseif_clause_list => L!("elseif_clause_list"),
Type::else_clause => L!("else_clause"),
Type::if_statement => L!("if_statement"),
Type::case_item => L!("case_item"),
Type::switch_statement => L!("switch_statement"),
Type::decorated_statement => L!("decorated_statement"),
Type::not_statement => L!("not_statement"),
Type::job_continuation => L!("job_continuation"),
Type::job_continuation_list => L!("job_continuation_list"),
Type::job_conjunction_continuation => L!("job_conjunction_continuation"),
Type::andor_job => L!("andor_job"),
Type::andor_job_list => L!("andor_job_list"),
Type::freestanding_argument_list => L!("freestanding_argument_list"),
Type::token_conjunction => L!("token_conjunction"),
Type::job_conjunction_continuation_list => L!("job_conjunction_continuation_list"),
Type::maybe_newlines => L!("maybe_newlines"),
Type::token_pipe => L!("token_pipe"),
Type::case_item_list => L!("case_item_list"),
Type::argument => L!("argument"),
Type::argument_list => L!("argument_list"),
Type::job_list => L!("job_list"),
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
}
// A way to visit nodes iteratively.
// This is pre-order. Each node is visited before its children.
// Example:
// let tv = Traversal::new(start);
// while let Some(node) = tv.next() {...}
pub struct Traversal<'a> {
stack: Vec<&'a dyn Node>,
}
impl<'a> Traversal<'a> {
// Construct starting with a node
pub fn new(n: &'a dyn Node) -> Self {
Self { stack: vec![n] }
}
}
impl<'a> Iterator for Traversal<'a> {
type Item = &'a dyn Node;
fn next(&mut self) -> Option<&'a dyn Node> {
let Some(node) = self.stack.pop() else {
return None;
};
// We want to visit in reverse order so the first child ends up on top of the stack.
node.accept(self, true /* reverse */);
Some(node)
}
}
impl<'a, 'v: 'a> NodeVisitor<'v> for Traversal<'a> {
fn visit(&mut self, node: &'a dyn Node) {
self.stack.push(node)
}
}
pub type SourceRangeList = Vec<SourceRange>;
/// Extra source ranges.
/// These are only generated if the corresponding flags are set.
#[derive(Default)]
pub struct Extras {
/// Set of comments, sorted by offset.
pub comments: SourceRangeList,
/// Set of semicolons, sorted by offset.
pub semis: SourceRangeList,
/// Set of error ranges, sorted by offset.
pub errors: SourceRangeList,
}
/// The ast type itself.
pub struct Ast {
// The top node.
// Its type depends on what was requested to parse.
top: Box<dyn NodeMut>,
/// Whether any errors were encountered during parsing.
pub any_error: bool,
/// Extra fields.
pub extras: Extras,
}
#[allow(clippy::derivable_impls)] // false positive
impl Default for Ast {
fn default() -> Ast {
Self {
top: Box::<String_>::default(),
any_error: false,
extras: Extras::default(),
}
}
}
impl Ast {
/// Construct an ast by parsing \p src as a job list.
/// The ast attempts to produce \p type as the result.
/// \p type may only be JobList or FreestandingArgumentList.
pub fn parse(
src: &wstr,
flags: ParseTreeFlags,
out_errors: Option<&mut ParseErrorList>,
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
) -> Self {
parse_from_top(src, flags, out_errors, Type::job_list)
}
/// Like parse(), but constructs a freestanding_argument_list.
pub fn parse_argument_list(
src: &wstr,
flags: ParseTreeFlags,
out_errors: Option<&mut ParseErrorList>,
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
) -> Self {
parse_from_top(src, flags, out_errors, Type::freestanding_argument_list)
}
/// \return a traversal, allowing iteration over the nodes.
pub fn walk(&'_ self) -> Traversal<'_> {
Traversal::new(self.top.as_node())
}
/// \return the top node. This has the type requested in the 'parse' method.
pub fn top(&self) -> &dyn Node {
self.top.as_node()
}
fn top_mut(&mut self) -> &mut dyn NodeMut {
&mut *self.top
}
/// \return whether any errors were encountered during parsing.
pub fn errored(&self) -> bool {
self.any_error
}
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
/// \return a textual representation of the tree.
/// Pass the original source as \p orig.
pub fn dump(&self, orig: &wstr) -> WString {
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
let mut result = WString::new();
for node in self.walk() {
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
let depth = get_depth(node);
// dot-| padding
result += &str::repeat("! ", depth)[..];
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
if let Some(n) = node.as_argument() {
result += "argument";
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
if let Some(argsrc) = n.try_source(orig) {
sprintf!(=> &mut result, ": '%ls'", argsrc);
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
} else if let Some(n) = node.as_keyword() {
sprintf!(=> &mut result, "keyword: %ls", n.keyword().to_wstr());
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
} else if let Some(n) = node.as_token() {
let desc = match n.token_type() {
ParseTokenType::string => {
let mut desc = WString::from_str("string");
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
if let Some(strsource) = n.try_source(orig) {
sprintf!(=> &mut desc, ": '%ls'", strsource);
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
desc
}
ParseTokenType::redirection => {
let mut desc = WString::from_str("redirection");
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
if let Some(strsource) = n.try_source(orig) {
sprintf!(=> &mut desc, ": '%ls'", strsource);
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
desc
}
ParseTokenType::end => WString::from_str("<;>"),
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
ParseTokenType::invalid => {
// This may occur with errors, e.g. we expected to see a string but saw a
// redirection.
WString::from_str("<error>")
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
_ => {
token_type_user_presentable_description(n.token_type(), ParseKeyword::none)
}
};
result += &desc[..];
} else {
result += &node.describe()[..];
}
result += "\n";
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
result
}
}
// \return the depth of a node, i.e. number of parent links.
fn get_depth(node: &dyn Node) -> usize {
let mut result = 0;
let mut cursor = node;
while let Some(parent) = cursor.parent() {
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
result += 1;
cursor = parent;
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
result
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
struct SourceRangeVisitor {
/// Total range we have encountered.
total: SourceRange,
/// Whether any node was found to be unsourced.
any_unsourced: bool,
}
impl<'a> NodeVisitor<'a> for SourceRangeVisitor {
fn visit(&mut self, node: &'a dyn Node) {
match node.category() {
Category::leaf => match node.as_leaf().unwrap().range() {
None => self.any_unsourced = true,
// Union with our range.
Some(range) if range.length > 0 => {
if self.total.length == 0 {
self.total = range;
} else {
let end =
(self.total.start + self.total.length).max(range.start + range.length);
self.total.start = self.total.start.min(range.start);
self.total.length = end - self.total.start;
}
}
_ => (),
},
_ => {
// Other node types recurse.
node.accept(self, false);
}
}
}
}
/// A token stream generates a sequence of parser tokens, permitting arbitrary lookahead.
struct TokenStream<'a> {
// We implement a queue with a simple circular buffer.
// Note that peek() returns an address, so we must not move elements which are peek'd.
// This prevents using vector (which may reallocate).
// Deque would work but is too heavyweight for just 2 items.
lookahead: [ParseToken; TokenStream::MAX_LOOKAHEAD],
// Starting index in our lookahead.
// The "first" token is at this index.
start: usize,
// Number of items in our lookahead.
count: usize,
// A reference to the original source.
src: &'a wstr,
// The tokenizer to generate new tokens.
tok: Tokenizer,
/// Any comment nodes are collected here.
/// These are only collected if parse_flag_include_comments is set.
comment_ranges: SourceRangeList,
}
impl<'a> TokenStream<'a> {
// The maximum number of lookahead supported.
const MAX_LOOKAHEAD: usize = 2;
fn new(src: &'a wstr, flags: ParseTreeFlags) -> Self {
Self {
lookahead: [ParseToken::new(ParseTokenType::invalid); Self::MAX_LOOKAHEAD],
start: 0,
count: 0,
src,
tok: Tokenizer::new(src, TokFlags::from(flags)),
comment_ranges: vec![],
}
}
/// \return the token at the given index, without popping it. If the token stream is exhausted,
/// it will have parse_token_type_t::terminate. idx = 0 means the next token, idx = 1 means the
/// next-next token, and so forth.
/// We must have that idx < kMaxLookahead.
fn peek(&mut self, idx: usize) -> &ParseToken {
assert!(idx < Self::MAX_LOOKAHEAD, "Trying to look too far ahead");
while idx >= self.count {
self.lookahead[Self::mask(self.start + self.count)] = self.next_from_tok();
self.count += 1
}
&self.lookahead[Self::mask(self.start + idx)]
}
/// Pop the next token.
fn pop(&mut self) -> ParseToken {
if self.count == 0 {
return self.next_from_tok();
}
let result = self.lookahead[self.start];
self.start = Self::mask(self.start + 1);
self.count -= 1;
result
}
// Helper to mask our circular buffer.
fn mask(idx: usize) -> usize {
idx % Self::MAX_LOOKAHEAD
}
/// \return the next parse token from the tokenizer.
/// This consumes and stores comments.
fn next_from_tok(&mut self) -> ParseToken {
loop {
let res = self.advance_1();
if res.typ == ParseTokenType::comment {
self.comment_ranges.push(res.range());
continue;
}
return res;
}
}
/// \return a new parse token, advancing the tokenizer.
/// This returns comments.
fn advance_1(&mut self) -> ParseToken {
let Some(token) = self.tok.next() else {
return ParseToken::new(ParseTokenType::terminate);
};
// Set the type, keyword, and whether there's a dash prefix. Note that this is quite
// sketchy, because it ignores quotes. This is the historical behavior. For example,
// `builtin --names` lists builtins, but `builtin "--names"` attempts to run --names as a
// command. Amazingly as of this writing (10/12/13) nobody seems to have noticed this.
// Squint at it really hard and it even starts to look like a feature.
let mut result = ParseToken::new(ParseTokenType::from(token.type_));
let text = self.tok.text_of(&token);
result.keyword = keyword_for_token(token.type_, text);
result.has_dash_prefix = text.starts_with('-');
result.is_help_argument = [L!("-h"), L!("--help")].contains(&text);
result.is_newline = result.typ == ParseTokenType::end && text == L!("\n");
result.may_be_variable_assignment = variable_assignment_equals_pos(text).is_some();
result.tok_error = token.error;
assert!(token.offset() < SOURCE_OFFSET_INVALID);
result.set_source_start(token.offset());
result.set_source_length(token.length());
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
if token.error != TokenizerError::none {
let subtoken_offset = token.error_offset_within_token();
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
// Skip invalid tokens that have a zero length, especially if they are at EOF.
if subtoken_offset < result.source_length() {
result.set_source_start(result.source_start() + subtoken_offset);
result.set_source_length(token.error_length());
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
}
result
}
}
/// This indicates a bug in fish code.
macro_rules! internal_error {
(
$self:ident,
$func:ident,
$fmt:expr
$(, $args:expr)*
$(,)?
) => {
FLOG!(
debug,
concat!(
"Internal parse error from {$func} - this indicates a bug in fish.",
$fmt,
)
$(, $args)*
);
FLOGF!(debug, "Encountered while parsing:<<<<\n{}\n>>>", $self.tokens.src);
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
panic!();
};
}
/// Report an error based on \p fmt for the tokens' range
macro_rules! parse_error {
(
$self:ident,
$token:expr,
$code:expr,
$fmt:expr
$(, $args:expr)*
$(,)?
) => {
let range = $token.range();
parse_error_range!($self, range, $code, $fmt $(, $args)*);
}
}
/// Report an error based on \p fmt for the source range \p range.
macro_rules! parse_error_range {
(
$self:ident,
$range:expr,
$code:expr,
$fmt:expr
$(, $args:expr)*
$(,)?
) => {
let text = if $self.out_errors.is_some() && !$self.unwinding {
Some(wgettext_maybe_fmt!($fmt $(, $args)*))
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
} else {
None
};
$self.any_error = true;
// Ignore additional parse errors while unwinding.
// These may come about e.g. from `true | and`.
if !$self.unwinding {
$self.unwinding = true;
FLOGF!(ast_construction, "%*sparse error - begin unwinding", $self.spaces(), "");
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
// TODO: can store this conditionally dependent on flags.
if $range.start() != SOURCE_OFFSET_INVALID {
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
$self.errors.push($range);
}
if let Some(errors) = &mut $self.out_errors {
let mut err = ParseError::default();
err.text = text.unwrap();
err.code = $code;
err.source_start = $range.start();
err.source_length = $range.length();
errors.push(err);
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
}
}
}
struct Populator<'a> {
/// Flags controlling parsing.
flags: ParseTreeFlags,
/// Set of semicolons, sorted by offset.
semis: SourceRangeList,
/// Set of error ranges, sorted by offset.
errors: SourceRangeList,
/// Stream of tokens which we consume.
tokens: TokenStream<'a>,
/** The type which we are attempting to parse, typically job_list but may be
freestanding_argument_list. */
top_type: Type,
/// If set, we are unwinding due to error recovery.
unwinding: bool,
/// If set, we have encountered an error.
any_error: bool,
/// The number of parent links of the node we are visiting
depth: usize,
// If non-null, populate with errors.
out_errors: Option<&'a mut ParseErrorList>,
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
impl<'s> NodeVisitorMut for Populator<'s> {
fn visit_mut(&mut self, node: &mut dyn NodeMut) -> VisitResult {
match node.typ() {
Type::argument => {
self.visit_argument(node.as_mut_argument().unwrap());
return VisitResult::Continue(());
}
Type::variable_assignment => {
self.visit_variable_assignment(node.as_mut_variable_assignment().unwrap());
return VisitResult::Continue(());
}
Type::job_continuation => {
self.visit_job_continuation(node.as_mut_job_continuation().unwrap());
return VisitResult::Continue(());
}
Type::token_base => {
self.visit_token(node.as_mut_token().unwrap());
return VisitResult::Continue(());
}
Type::keyword_base => {
return self.visit_keyword(node.as_mut_keyword().unwrap());
}
Type::maybe_newlines => {
self.visit_maybe_newlines(node.as_mut_maybe_newlines().unwrap());
return VisitResult::Continue(());
}
_ => (),
}
match node.category() {
Category::leaf => {}
// Visit branch nodes by just calling accept() to visit their fields.
Category::branch => {
// This field is a direct embedding of an AST value.
node.accept_mut(self, false);
return VisitResult::Continue(());
}
Category::list => {
// This field is an embedding of an array of (pointers to) ContentsNode.
// Parse as many as we can.
match node.typ() {
Type::andor_job_list => self.populate_list::<AndorJobList>(
node.as_mut_andor_job_list().unwrap(),
false,
),
Type::argument_list => self
.populate_list::<ArgumentList>(node.as_mut_argument_list().unwrap(), false),
Type::argument_or_redirection_list => self
.populate_list::<ArgumentOrRedirectionList>(
node.as_mut_argument_or_redirection_list().unwrap(),
false,
),
Type::case_item_list => self.populate_list::<CaseItemList>(
node.as_mut_case_item_list().unwrap(),
false,
),
Type::elseif_clause_list => self.populate_list::<ElseifClauseList>(
node.as_mut_elseif_clause_list().unwrap(),
false,
),
Type::job_conjunction_continuation_list => self
.populate_list::<JobConjunctionContinuationList>(
node.as_mut_job_conjunction_continuation_list().unwrap(),
false,
),
Type::job_continuation_list => self.populate_list::<JobContinuationList>(
node.as_mut_job_continuation_list().unwrap(),
false,
),
Type::job_list => {
self.populate_list::<JobList>(node.as_mut_job_list().unwrap(), false)
}
Type::variable_assignment_list => self.populate_list::<VariableAssignmentList>(
node.as_mut_variable_assignment_list().unwrap(),
false,
),
_ => (),
}
}
}
VisitResult::Continue(())
}
fn will_visit_fields_of(&mut self, node: &mut dyn NodeMut) {
FLOGF!(
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
ast_construction,
"%*swill_visit %ls",
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
self.spaces(),
"",
node.describe()
);
self.depth += 1
}
fn did_visit_fields_of<'a>(&'a mut self, node: &'a dyn NodeMut, flow: VisitResult) {
self.depth -= 1;
if self.unwinding {
return;
}
let VisitResult::Break(error) = flow else {
return;
};
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
// We believe the node is some sort of block statement. Attempt to find a source range
// for the block's keyword (for, if, etc) and a user-presentable description. This
// is used to provide better error messages. Note at this point the parse tree is
// incomplete; in particular parent nodes are not set.
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
let mut cursor = node;
let header = loop {
match cursor.typ() {
Type::block_statement => {
cursor = cursor.as_block_statement().unwrap().header.embedded_node();
}
Type::for_header => {
let n = cursor.as_for_header().unwrap();
break Some((n.kw_for.range.unwrap(), L!("for loop")));
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
Type::while_header => {
let n = cursor.as_while_header().unwrap();
break Some((n.kw_while.range.unwrap(), L!("while loop")));
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
Type::function_header => {
let n = cursor.as_function_header().unwrap();
break Some((n.kw_function.range.unwrap(), L!("function definition")));
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
Type::begin_header => {
let n = cursor.as_begin_header().unwrap();
break Some((n.kw_begin.range.unwrap(), L!("begin")));
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
Type::if_statement => {
let n = cursor.as_if_statement().unwrap();
break Some((n.if_clause.kw_if.range.unwrap(), L!("if statement")));
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
Type::switch_statement => {
let n = cursor.as_switch_statement().unwrap();
break Some((n.kw_switch.range.unwrap(), L!("switch statement")));
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
_ => break None,
}
};
if let Some((header_kw_range, enclosing_stmt)) = header {
parse_error_range!(
self,
header_kw_range,
ParseErrorCode::generic,
"Missing end to balance this %ls",
enclosing_stmt
);
} else {
parse_error!(
self,
error.token,
ParseErrorCode::generic,
"Expected %ls, but found %ls",
keywords_user_presentable_description(error.allowed_keywords),
error.token.user_presentable_description(),
);
}
}
// We currently only have a handful of union pointer types.
// Handle them directly.
fn visit_argument_or_redirection(
&mut self,
node: &mut Box<ArgumentOrRedirectionVariant>,
) -> VisitResult {
if let Some(arg) = self.try_parse::<Argument>() {
**node = ArgumentOrRedirectionVariant::Argument(*arg);
} else if let Some(redir) = self.try_parse::<Redirection>() {
**node = ArgumentOrRedirectionVariant::Redirection(*redir);
} else {
internal_error!(
self,
visit_argument_or_redirection,
"Unable to parse argument or redirection"
);
}
VisitResult::Continue(())
}
fn visit_block_statement_header(
&mut self,
node: &mut Box<BlockStatementHeaderVariant>,
) -> VisitResult {
*node = self.allocate_populate_block_header();
VisitResult::Continue(())
}
fn visit_statement(&mut self, node: &mut Box<StatementVariant>) -> VisitResult {
*node = self.allocate_populate_statement_contents();
VisitResult::Continue(())
}
fn visit_decorated_statement_decorator(
&mut self,
node: &mut Option<DecoratedStatementDecorator>,
) {
*node = self.try_parse::<DecoratedStatementDecorator>().map(|b| *b);
}
fn visit_job_conjunction_decorator(&mut self, node: &mut Option<JobConjunctionDecorator>) {
*node = self.try_parse::<JobConjunctionDecorator>().map(|b| *b);
}
fn visit_else_clause(&mut self, node: &mut Option<ElseClause>) {
*node = self.try_parse::<ElseClause>().map(|b| *b);
}
fn visit_semi_nl(&mut self, node: &mut Option<SemiNl>) {
*node = self.try_parse::<SemiNl>().map(|b| *b);
}
fn visit_time(&mut self, node: &mut Option<KeywordTime>) {
*node = self.try_parse::<KeywordTime>().map(|b| *b);
}
fn visit_token_background(&mut self, node: &mut Option<TokenBackground>) {
*node = self.try_parse::<TokenBackground>().map(|b| *b);
}
}
/// Helper to describe a list of keywords.
/// TODO: these need to be localized properly.
fn keywords_user_presentable_description(kws: &'static [ParseKeyword]) -> WString {
assert!(!kws.is_empty(), "Should not be empty list");
if kws.len() == 1 {
return sprintf!(L!("keyword '%ls'"), kws[0]);
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
let mut res = L!("keywords ").to_owned();
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
for (i, kw) in kws.iter().enumerate() {
if i != 0 {
res += L!(" or ");
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
res += &sprintf!(L!("'%ls'"), *kw)[..];
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
res
}
/// Helper to describe a list of token types.
/// TODO: these need to be localized properly.
fn token_types_user_presentable_description(types: &'static [ParseTokenType]) -> WString {
assert!(!types.is_empty(), "Should not be empty list");
let mut res = WString::new();
for typ in types {
if !res.is_empty() {
res += L!(" or ");
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
res += &token_type_user_presentable_description(*typ, ParseKeyword::none)[..];
}
res
}
impl<'s> Populator<'s> {
/// Construct from a source, flags, top type, and out_errors, which may be null.
fn new(
src: &'s wstr,
flags: ParseTreeFlags,
top_type: Type,
out_errors: Option<&'s mut ParseErrorList>,
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
) -> Self {
Self {
flags,
semis: vec![],
errors: vec![],
tokens: TokenStream::new(src, flags),
top_type,
unwinding: false,
any_error: false,
depth: 0,
out_errors,
}
}
/// Helper for FLOGF. This returns a number of spaces appropriate for a '%*c' format.
fn spaces(&self) -> usize {
self.depth * 2
}
/// \return the parser's status.
fn status(&mut self) -> ParserStatus {
if self.unwinding {
ParserStatus::unwinding
} else if self.flags.contains(ParseTreeFlags::LEAVE_UNTERMINATED)
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
&& self.peek_type(0) == ParseTokenType::terminate
{
ParserStatus::unsourcing
} else {
ParserStatus::ok
}
}
/// \return whether any leaf nodes we visit should be marked as unsourced.
fn unsource_leaves(&mut self) -> bool {
matches!(
self.status(),
ParserStatus::unsourcing | ParserStatus::unwinding
)
}
/// \return whether we permit an incomplete parse tree.
fn allow_incomplete(&self) -> bool {
self.flags.contains(ParseTreeFlags::LEAVE_UNTERMINATED)
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
/// \return whether a list type \p type allows arbitrary newlines in it.
fn list_type_chomps_newlines(&self, typ: Type) -> bool {
match typ {
Type::argument_list => {
// Hackish. If we are producing a freestanding argument list, then it allows
// semicolons, for hysterical raisins.
self.top_type == Type::freestanding_argument_list
}
Type::argument_or_redirection_list => {
// No newlines inside arguments.
false
}
Type::variable_assignment_list => {
// No newlines inside variable assignment lists.
false
}
Type::job_list => {
// Like echo a \n \n echo b
true
}
Type::case_item_list => {
// Like switch foo \n \n \n case a \n end
true
}
Type::andor_job_list => {
// Like while true ; \n \n and true ; end
true
}
Type::elseif_clause_list => {
// Like if true ; \n \n else if false; end
true
}
Type::job_conjunction_continuation_list => {
// This would be like echo a && echo b \n && echo c
// We could conceivably support this but do not now.
false
}
Type::job_continuation_list => {
// This would be like echo a \n | echo b
// We could conceivably support this but do not now.
false
}
_ => {
internal_error!(
self,
list_type_chomps_newlines,
"Type %ls not handled",
ast_type_to_string(typ)
);
}
}
}
/// \return whether a list type \p type allows arbitrary semicolons in it.
fn list_type_chomps_semis(&self, typ: Type) -> bool {
match typ {
Type::argument_list => {
// Hackish. If we are producing a freestanding argument list, then it allows
// semicolons, for hysterical raisins.
// That is, this is OK: complete -c foo -a 'x ; y ; z'
// But this is not: foo x ; y ; z
self.top_type == Type::freestanding_argument_list
}
Type::argument_or_redirection_list | Type::variable_assignment_list => false,
Type::job_list => {
// Like echo a ; ; echo b
true
}
Type::case_item_list => {
// Like switch foo ; ; ; case a \n end
// This is historically allowed.
true
}
Type::andor_job_list => {
// Like while true ; ; ; and true ; end
true
}
Type::elseif_clause_list => {
// Like if true ; ; ; else if false; end
false
}
Type::job_conjunction_continuation_list => {
// Like echo a ; ; && echo b. Not supported.
false
}
Type::job_continuation_list => {
// This would be like echo a ; | echo b
// Not supported.
// We could conceivably support this but do not now.
false
}
_ => {
internal_error!(
self,
list_type_chomps_semis,
"Type %ls not handled",
ast_type_to_string(typ)
);
}
}
}
/// Chomp extra comments, semicolons, etc. for a given list type.
fn chomp_extras(&mut self, typ: Type) {
let chomp_semis = self.list_type_chomps_semis(typ);
let chomp_newlines = self.list_type_chomps_newlines(typ);
loop {
let peek = self.tokens.peek(0);
if chomp_newlines && peek.typ == ParseTokenType::end && peek.is_newline {
// Just skip this newline, no need to save it.
self.tokens.pop();
} else if chomp_semis && peek.typ == ParseTokenType::end && !peek.is_newline {
let tok = self.tokens.pop();
// Perhaps save this extra semi.
if self.flags.contains(ParseTreeFlags::SHOW_EXTRA_SEMIS) {
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
self.semis.push(tok.range());
}
} else {
break;
}
}
}
/// \return whether a list type should recover from errors.s
/// That is, whether we should stop unwinding when we encounter this type.
fn list_type_stops_unwind(&self, typ: Type) -> bool {
typ == Type::job_list && self.flags.contains(ParseTreeFlags::CONTINUE_AFTER_ERROR)
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
/// \return a reference to a non-comment token at index \p idx.
fn peek_token(&mut self, idx: usize) -> &ParseToken {
self.tokens.peek(idx)
}
/// \return the type of a non-comment token.
fn peek_type(&mut self, idx: usize) -> ParseTokenType {
self.peek_token(idx).typ
}
/// Consume the next token, chomping any comments.
/// It is an error to call this unless we know there is a non-terminate token available.
/// \return the token.
fn consume_any_token(&mut self) -> ParseToken {
let tok = self.tokens.pop();
assert!(
tok.typ != ParseTokenType::comment,
"Should not be a comment"
);
assert!(
tok.typ != ParseTokenType::terminate,
"Cannot consume terminate token, caller should check status first"
);
tok
}
/// Consume the next token which is expected to be of the given type.
fn consume_token_type(&mut self, typ: ParseTokenType) -> SourceRange {
assert!(
typ != ParseTokenType::terminate,
"Should not attempt to consume terminate token"
);
let tok = self.consume_any_token();
if tok.typ != typ {
parse_error!(
self,
tok,
ParseErrorCode::generic,
"Expected %ls, but found %ls",
token_type_user_presentable_description(typ, ParseKeyword::none),
tok.user_presentable_description()
);
return SourceRange::new(0, 0);
}
tok.range()
}
/// The next token could not be parsed at the top level.
/// For example a trailing end like `begin ; end ; end`
/// Or an unexpected redirection like `>`
/// Consume it and add an error.
fn consume_excess_token_generating_error(&mut self) {
let tok = self.consume_any_token();
// In the rare case that we are parsing a freestanding argument list and not a job list,
// generate a generic error.
// TODO: this is a crummy message if we get a tokenizer error, for example:
// complete -c foo -a "'abc"
if self.top_type == Type::freestanding_argument_list {
parse_error!(
self,
tok,
ParseErrorCode::generic,
"Expected %ls, but found %ls",
token_type_user_presentable_description(ParseTokenType::string, ParseKeyword::none),
tok.user_presentable_description()
);
return;
}
assert!(self.top_type == Type::job_list);
match tok.typ {
ParseTokenType::string => {
// There are three keywords which end a job list.
match tok.keyword {
ParseKeyword::kw_end => {
parse_error!(
self,
tok,
ParseErrorCode::unbalancing_end,
"'end' outside of a block"
);
}
ParseKeyword::kw_else => {
parse_error!(
self,
tok,
ParseErrorCode::unbalancing_else,
"'else' builtin not inside of if block"
);
}
ParseKeyword::kw_case => {
parse_error!(
self,
tok,
ParseErrorCode::unbalancing_case,
"'case' builtin not inside of switch block"
);
}
_ => {
internal_error!(
self,
consume_excess_token_generating_error,
"Token %ls should not have prevented parsing a job list",
tok.user_presentable_description()
);
}
}
}
ParseTokenType::redirection if self.peek_type(0) == ParseTokenType::string => {
let next = self.tokens.pop();
parse_error_range!(
self,
next.range().combine(tok.range()),
ParseErrorCode::generic,
"Expected a string, but found a redirection"
);
}
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
ParseTokenType::pipe
| ParseTokenType::redirection
| ParseTokenType::background
| ParseTokenType::andand
| ParseTokenType::oror => {
parse_error!(
self,
tok,
ParseErrorCode::generic,
"Expected a string, but found %ls",
tok.user_presentable_description()
);
}
ParseTokenType::tokenizer_error => {
parse_error!(
self,
tok,
ParseErrorCode::from(tok.tok_error),
"%ls",
tok.tok_error
);
}
ParseTokenType::end => {
internal_error!(
self,
consume_excess_token_generating_error,
"End token should never be excess"
);
}
ParseTokenType::terminate => {
internal_error!(
self,
consume_excess_token_generating_error,
"Terminate token should never be excess"
);
}
_ => {
internal_error!(
self,
consume_excess_token_generating_error,
"Unexpected excess token type: %ls",
tok.user_presentable_description()
);
}
}
}
/// This is for optional values and for lists.
/// A true return means we should descend into the production, false means stop.
/// Note that the argument is always nullptr and should be ignored. It is provided strictly
/// for overloading purposes.
fn can_parse(&mut self, node: &dyn Node) -> bool {
match node.typ() {
Type::job_conjunction => {
let token = self.peek_token(0);
if token.typ != ParseTokenType::string {
return false;
}
!matches!(
token.keyword,
// These end a job list.
ParseKeyword::kw_end | ParseKeyword::kw_else | ParseKeyword::kw_case
)
}
Type::argument => self.peek_type(0) == ParseTokenType::string,
Type::redirection => self.peek_type(0) == ParseTokenType::redirection,
Type::argument_or_redirection => {
[ParseTokenType::string, ParseTokenType::redirection].contains(&self.peek_type(0))
}
Type::variable_assignment => {
// Do we have a variable assignment at all?
if !self.peek_token(0).may_be_variable_assignment {
return false;
}
// What is the token after it?
match self.peek_type(1) {
ParseTokenType::string => {
// We have `a= cmd` and should treat it as a variable assignment.
true
}
ParseTokenType::terminate => {
// We have `a=` which is OK if we are allowing incomplete, an error
// otherwise.
self.allow_incomplete()
}
_ => {
// We have e.g. `a= >` which is an error.
// Note that we do not produce an error here. Instead we return false
// so this the token will be seen by allocate_populate_statement_contents.
false
}
}
}
Type::token_base => node
.as_token()
.unwrap()
.allows_token(self.peek_token(0).typ),
// Note we have specific overloads for our keyword nodes, as they need custom logic.
Type::keyword_base => {
let keyword = node.as_keyword().unwrap();
match keyword.allowed_keywords() {
// job conjunction decorator
[ParseKeyword::kw_and, ParseKeyword::kw_or] => {
// This is for a job conjunction like `and stuff`
// But if it's `and --help` then we treat it as an ordinary command.
keyword.allows_keyword(self.peek_token(0).keyword)
&& !self.peek_token(1).is_help_argument
}
// decorated statement decorator
[ParseKeyword::kw_command, ParseKeyword::kw_builtin, ParseKeyword::kw_exec] => {
// Here the keyword is 'command' or 'builtin' or 'exec'.
// `command stuff` executes a command called stuff.
// `command -n` passes the -n argument to the 'command' builtin.
// `command` by itself is a command.
if !keyword.allows_keyword(self.peek_token(0).keyword) {
return false;
}
let tok1 = self.peek_token(1);
tok1.typ == ParseTokenType::string && !tok1.is_dash_prefix_string()
}
[ParseKeyword::kw_time] => {
// Time keyword is only the time builtin if the next argument doesn't
// have a dash.
keyword.allows_keyword(self.peek_token(0).keyword)
&& !self.peek_token(1).is_dash_prefix_string()
}
_ => panic!("Unexpected keyword in can_parse()"),
}
}
Type::job_continuation => self.peek_type(0) == ParseTokenType::pipe,
Type::job_conjunction_continuation => {
[ParseTokenType::andand, ParseTokenType::oror].contains(&self.peek_type(0))
}
Type::andor_job => {
match self.peek_token(0).keyword {
ParseKeyword::kw_and | ParseKeyword::kw_or => {
// Check that the argument to and/or is a string that's not help. Otherwise
// it's either 'and --help' or a naked 'and', and not part of this list.
let nexttok = self.peek_token(1);
nexttok.typ == ParseTokenType::string && !nexttok.is_help_argument
}
_ => false,
}
}
Type::elseif_clause => {
self.peek_token(0).keyword == ParseKeyword::kw_else
&& self.peek_token(1).keyword == ParseKeyword::kw_if
}
Type::else_clause => self.peek_token(0).keyword == ParseKeyword::kw_else,
Type::case_item => self.peek_token(0).keyword == ParseKeyword::kw_case,
_ => panic!("Unexpected token type in can_parse()"),
}
}
/// Given that we are a list of type ListNodeType, whose contents type is ContentsNode,
/// populate as many elements as we can.
/// If exhaust_stream is set, then keep going until we get parse_token_type_t::terminate.
fn populate_list<ListType: List>(&mut self, list: &mut ListType, exhaust_stream: bool)
where
<ListType as List>::ContentsNode: NodeMut,
{
assert!(list.contents().is_empty(), "List is not initially empty");
// Do not attempt to parse a list if we are unwinding.
if self.unwinding {
assert!(
!exhaust_stream,
"exhaust_stream should only be set at top level, and so we should not be unwinding"
);
// Mark in the list that it was unwound.
FLOGF!(
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
ast_construction,
"%*sunwinding %ls",
self.spaces(),
"",
ast_type_to_string(list.typ())
);
assert!(list.contents().is_empty(), "Should be an empty list");
return;
}
// We're going to populate a vector with our nodes.
// Later on we will copy this to the heap with a single allocation.
let mut contents = vec![];
loop {
// If we are unwinding, then either we recover or we break the loop, dependent on the
// loop type.
if self.unwinding {
if !self.list_type_stops_unwind(list.typ()) {
break;
}
// We are going to stop unwinding.
// Rather hackish. Just chomp until we get to a string or end node.
loop {
let typ = self.peek_type(0);
if [
ParseTokenType::string,
ParseTokenType::terminate,
ParseTokenType::end,
]
.contains(&typ)
{
break;
}
let tok = self.tokens.pop();
self.errors.push(tok.range());
FLOGF!(
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
ast_construction,
"%*schomping range %u-%u",
self.spaces(),
"",
tok.source_start(),
tok.source_length()
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
);
}
FLOGF!(ast_construction, "%*sdone unwinding", self.spaces(), "");
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
self.unwinding = false;
}
// Chomp semis and newlines.
self.chomp_extras(list.typ());
// Now try parsing a node.
if let Some(node) = self.try_parse::<ListType::ContentsNode>() {
// #7201: Minimize reallocations of contents vector
if contents.is_empty() {
contents.reserve(64);
}
contents.push(node);
} else if exhaust_stream && self.peek_type(0) != ParseTokenType::terminate {
// We aren't allowed to stop. Produce an error and keep going.
self.consume_excess_token_generating_error()
} else {
// We either stop once we can't parse any more of this contents node, or we
// exhausted the stream as requested.
break;
}
}
// Populate our list from our contents.
if !contents.is_empty() {
assert!(
contents.len() <= u32::MAX.try_into().unwrap(),
"Contents size out of bounds"
);
assert!(list.contents().is_empty(), "List should still be empty");
*list.contents_mut() = contents;
}
FLOGF!(
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
ast_construction,
"%*s%ls size: %lu",
self.spaces(),
"",
ast_type_to_string(list.typ()),
list.count()
);
}
/// Allocate and populate a statement contents pointer.
/// This must never return null.
fn allocate_populate_statement_contents(&mut self) -> Box<StatementVariant> {
// In case we get a parse error, we still need to return something non-null. Use a
// decorated statement; all of its leaf nodes will end up unsourced.
fn got_error(slf: &mut Populator<'_>) -> Box<StatementVariant> {
assert!(slf.unwinding, "Should have produced an error");
new_decorated_statement(slf)
}
fn new_decorated_statement(slf: &mut Populator<'_>) -> Box<StatementVariant> {
let embedded = slf.allocate_visit::<DecoratedStatement>();
Box::new(StatementVariant::DecoratedStatement(*embedded))
}
if self.peek_token(0).typ == ParseTokenType::terminate && self.allow_incomplete() {
// This may happen if we just have a 'time' prefix.
// Construct a decorated statement, which will be unsourced.
self.allocate_visit::<DecoratedStatement>();
} else if self.peek_token(0).typ != ParseTokenType::string {
// We may be unwinding already; do not produce another error.
// For example in `true | and`.
parse_error!(
self,
self.peek_token(0),
ParseErrorCode::generic,
"Expected a command, but found %ls",
self.peek_token(0).user_presentable_description()
);
return got_error(self);
} else if self.peek_token(0).may_be_variable_assignment {
// Here we have a variable assignment which we chose to not parse as a variable
// assignment because there was no string after it.
// Ensure we consume the token, so we don't get back here again at the same place.
let token = &self.consume_any_token();
let text = &self.tokens.src
[token.source_start()..token.source_start() + token.source_length()];
let equals_pos = variable_assignment_equals_pos(text).unwrap();
let variable = &text[..equals_pos];
let value = &text[equals_pos + 1..];
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
parse_error!(
self,
token,
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
ParseErrorCode::bare_variable_assignment,
ERROR_BAD_COMMAND_ASSIGN_ERR_MSG,
variable,
value
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
);
return got_error(self);
}
// In some cases a block starter is a decorated statement instead, mostly if invoked with "--help".
// The logic here is subtle:
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
//
// If we are 'begin', it's only really a block if it has no arguments.
// If we are 'function' or another block starter, then we are a non-block if we are invoked with -h or --help
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
// If we are anything else, we require an argument, so do the same thing if the subsequent
// token is a statement terminator.
if self.peek_token(0).typ == ParseTokenType::string {
// If we are one of these, then look for specifically help arguments. Otherwise, if the next token
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
// looks like an option (starts with a dash), then parse it as a decorated statement.
let help_only_kws = [
ParseKeyword::kw_begin,
ParseKeyword::kw_function,
ParseKeyword::kw_if,
ParseKeyword::kw_switch,
ParseKeyword::kw_while,
];
if (help_only_kws.contains(&self.peek_token(0).keyword)
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
&& self.peek_token(1).is_help_argument)
|| (!help_only_kws.contains(&self.peek_token(0).keyword)
&& self.peek_token(1).is_dash_prefix_string())
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
{
return new_decorated_statement(self);
}
// Likewise if the next token doesn't look like an argument at all. This corresponds to
// e.g. a "naked if".
let naked_invocation_invokes_help = ![ParseKeyword::kw_begin, ParseKeyword::kw_end]
.contains(&self.peek_token(0).keyword);
if naked_invocation_invokes_help
&& [ParseTokenType::end, ParseTokenType::terminate]
.contains(&self.peek_token(1).typ)
{
return new_decorated_statement(self);
}
}
match self.peek_token(0).keyword {
ParseKeyword::kw_not | ParseKeyword::kw_exclam => {
let embedded = self.allocate_visit::<NotStatement>();
Box::new(StatementVariant::NotStatement(*embedded))
}
ParseKeyword::kw_for
| ParseKeyword::kw_while
| ParseKeyword::kw_function
| ParseKeyword::kw_begin => {
let embedded = self.allocate_visit::<BlockStatement>();
Box::new(StatementVariant::BlockStatement(*embedded))
}
ParseKeyword::kw_if => {
let embedded = self.allocate_visit::<IfStatement>();
Box::new(StatementVariant::IfStatement(*embedded))
}
ParseKeyword::kw_switch => {
let embedded = self.allocate_visit::<SwitchStatement>();
Box::new(StatementVariant::SwitchStatement(*embedded))
}
ParseKeyword::kw_end => {
// 'end' is forbidden as a command.
// For example, `if end` or `while end` will produce this error.
// We still have to descend into the decorated statement because
// we can't leave our pointer as null.
parse_error!(
self,
self.peek_token(0),
ParseErrorCode::generic,
"Expected a command, but found %ls",
self.peek_token(0).user_presentable_description()
);
return got_error(self);
}
_ => new_decorated_statement(self),
}
}
/// Allocate and populate a block statement header.
/// This must never return null.
fn allocate_populate_block_header(&mut self) -> Box<BlockStatementHeaderVariant> {
Box::new(match self.peek_token(0).keyword {
ParseKeyword::kw_for => {
let embedded = self.allocate_visit::<ForHeader>();
BlockStatementHeaderVariant::ForHeader(*embedded)
}
ParseKeyword::kw_while => {
let embedded = self.allocate_visit::<WhileHeader>();
BlockStatementHeaderVariant::WhileHeader(*embedded)
}
ParseKeyword::kw_function => {
let embedded = self.allocate_visit::<FunctionHeader>();
BlockStatementHeaderVariant::FunctionHeader(*embedded)
}
ParseKeyword::kw_begin => {
let embedded = self.allocate_visit::<BeginHeader>();
BlockStatementHeaderVariant::BeginHeader(*embedded)
}
_ => {
internal_error!(
self,
allocate_populate_block_header,
"should not have descended into block_header"
);
}
})
}
fn try_parse<T: NodeMut + Default>(&mut self) -> Option<Box<T>> {
// TODO Optimize this.
let prototype = T::default();
if !self.can_parse(&prototype) {
return None;
}
Some(self.allocate_visit())
}
/// Given a node type, allocate it and invoke its default constructor.
/// \return the resulting Node
fn allocate<T: NodeMut + Default>(&self) -> Box<T> {
let result = Box::<T>::default();
FLOGF!(
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
ast_construction,
"%*smake %ls %ls",
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
self.spaces(),
"",
ast_type_to_string(result.typ()),
format!("{result:p}")
);
result
}
// Given a node type, allocate it, invoke its default constructor,
// and then visit it as a field.
// \return the resulting Node pointer. It is never null.
fn allocate_visit<T: NodeMut + Default>(&mut self) -> Box<T> {
let mut result = Box::<T>::default();
self.visit_mut(&mut *result);
result
}
fn visit_argument(&mut self, arg: &mut Argument) {
if self.unsource_leaves() {
arg.range = None;
return;
}
arg.range = Some(self.consume_token_type(ParseTokenType::string));
}
fn visit_variable_assignment(&mut self, varas: &mut VariableAssignment) {
if self.unsource_leaves() {
varas.range = None;
return;
}
if !self.peek_token(0).may_be_variable_assignment {
internal_error!(
self,
visit_variable_assignment,
"Should not have created variable_assignment_t from this token"
);
}
varas.range = Some(self.consume_token_type(ParseTokenType::string));
}
fn visit_job_continuation(&mut self, node: &mut JobContinuation) {
// Special error handling to catch 'and' and 'or' in pipelines, like `true | and false`.
if [ParseKeyword::kw_and, ParseKeyword::kw_or].contains(&self.peek_token(1).keyword) {
parse_error!(
self,
self.peek_token(1),
ParseErrorCode::andor_in_pipeline,
INVALID_PIPELINE_CMD_ERR_MSG,
self.peek_token(1).keyword
);
}
node.accept_mut(self, false);
}
// Overload for token fields.
fn visit_token(&mut self, token: &mut dyn Token) {
if self.unsource_leaves() {
*token.range_mut() = None;
return;
}
if !token.allows_token(self.peek_token(0).typ) {
if self.flags.contains(ParseTreeFlags::LEAVE_UNTERMINATED)
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
&& [
TokenizerError::unterminated_quote,
TokenizerError::unterminated_subshell,
]
.contains(&self.peek_token(0).tok_error)
{
return;
}
parse_error!(
self,
self.peek_token(0),
ParseErrorCode::generic,
"Expected %ls, but found %ls",
token_types_user_presentable_description(token.allowed_tokens()),
self.peek_token(0).user_presentable_description()
);
*token.range_mut() = None;
return;
}
let tok = self.consume_any_token();
*token.token_type_mut() = tok.typ;
*token.range_mut() = Some(tok.range());
}
// Overload for keyword fields.
fn visit_keyword(&mut self, keyword: &mut dyn Keyword) -> VisitResult {
if self.unsource_leaves() {
*keyword.range_mut() = None;
return VisitResult::Continue(());
}
if !keyword.allows_keyword(self.peek_token(0).keyword) {
*keyword.range_mut() = None;
if self.flags.contains(ParseTreeFlags::LEAVE_UNTERMINATED)
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
&& [
TokenizerError::unterminated_quote,
TokenizerError::unterminated_subshell,
]
.contains(&self.peek_token(0).tok_error)
{
return VisitResult::Continue(());
}
// Special error reporting for keyword_t<kw_end>.
let allowed_keywords = keyword.allowed_keywords();
if keyword.allowed_keywords() == [ParseKeyword::kw_end] {
return VisitResult::Break(MissingEndError {
allowed_keywords,
token: *self.peek_token(0),
});
} else {
parse_error!(
self,
self.peek_token(0),
ParseErrorCode::generic,
"Expected %ls, but found %ls",
keywords_user_presentable_description(allowed_keywords),
self.peek_token(0).user_presentable_description(),
);
return VisitResult::Continue(());
}
}
let tok = self.consume_any_token();
*keyword.keyword_mut() = tok.keyword;
*keyword.range_mut() = Some(tok.range());
VisitResult::Continue(())
}
fn visit_maybe_newlines(&mut self, nls: &mut MaybeNewlines) {
if self.unsource_leaves() {
nls.range = None;
return;
}
let mut range = SourceRange::new(0, 0);
// TODO: it would be nice to have the start offset be the current position in the token
// stream, even if there are no newlines.
while self.peek_token(0).is_newline {
let r = self.consume_token_type(ParseTokenType::end);
if range.length == 0 {
range = r;
} else {
range.length = r.start + r.length - range.start
}
}
nls.range = Some(range);
}
}
/// The status of our parser.
enum ParserStatus {
/// Parsing is going just fine, thanks for asking.
ok,
/// We have exhausted the token stream, but the caller was OK with an incomplete parse tree.
/// All further leaf nodes should have the unsourced flag set.
unsourcing,
/// We encountered an parse error and are "unwinding."
/// Do not consume any tokens until we get back to a list type which stops unwinding.
unwinding,
}
fn parse_from_top(
src: &wstr,
flags: ParseTreeFlags,
out_errors: Option<&mut ParseErrorList>,
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
top_type: Type,
) -> Ast {
assert!(
[Type::job_list, Type::freestanding_argument_list].contains(&top_type),
"Invalid top type"
);
let mut ast = Ast::default();
let mut pops = Populator::new(src, flags, top_type, out_errors);
if top_type == Type::job_list {
let mut list = pops.allocate::<JobList>();
pops.populate_list(&mut *list, true /* exhaust_stream */);
ast.top = list;
} else {
let mut list = pops.allocate::<FreestandingArgumentList>();
pops.populate_list(&mut list.arguments, true /* exhaust_stream */);
ast.top = list;
}
// Chomp trailing extras, etc.
pops.chomp_extras(Type::job_list);
ast.any_error = pops.any_error;
ast.extras = Extras {
comments: pops.tokens.comment_ranges,
semis: pops.semis,
errors: pops.errors,
};
if top_type == Type::job_list {
// Set all parent nodes.
// It turns out to be more convenient to do this after the parse phase.
ast.top_mut()
.as_mut_job_list()
.as_mut()
.unwrap()
.set_parents();
} else {
ast.top_mut()
.as_mut_freestanding_argument_list()
.as_mut()
.unwrap()
.set_parents();
}
ast
}
/// \return tokenizer flags corresponding to parse tree flags.
impl From<ParseTreeFlags> for TokFlags {
fn from(flags: ParseTreeFlags) -> Self {
let mut tok_flags = TokFlags(0);
// Note we do not need to respect parse_flag_show_blank_lines, no clients are interested
// in them.
if flags.contains(ParseTreeFlags::INCLUDE_COMMENTS) {
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
tok_flags |= TOK_SHOW_COMMENTS;
}
if flags.contains(ParseTreeFlags::ACCEPT_INCOMPLETE_TOKENS) {
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
tok_flags |= TOK_ACCEPT_UNFINISHED;
}
if flags.contains(ParseTreeFlags::CONTINUE_AFTER_ERROR) {
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
tok_flags |= TOK_CONTINUE_AFTER_ERROR
}
tok_flags
}
}
/// Convert from Tokenizer's token type to a parse_token_t type.
impl From<TokenType> for ParseTokenType {
fn from(token_type: TokenType) -> Self {
match token_type {
TokenType::string => ParseTokenType::string,
TokenType::pipe => ParseTokenType::pipe,
TokenType::andand => ParseTokenType::andand,
TokenType::oror => ParseTokenType::oror,
TokenType::end => ParseTokenType::end,
TokenType::background => ParseTokenType::background,
TokenType::redirect => ParseTokenType::redirection,
TokenType::error => ParseTokenType::tokenizer_error,
TokenType::comment => ParseTokenType::comment,
}
}
}
fn is_keyword_char(c: char) -> bool {
('a'..='z').contains(&c)
|| ('A'..='Z').contains(&c)
|| ('0'..='9').contains(&c)
|| c == '\''
|| c == '"'
|| c == '\\'
|| c == '\n'
|| c == '!'
}
/// Given a token, returns the keyword it matches, or ParseKeyword::none.
fn keyword_for_token(tok: TokenType, token: &wstr) -> ParseKeyword {
/* Only strings can be keywords */
if tok != TokenType::string {
return ParseKeyword::none;
}
// If token is clean (which most are), we can compare it directly. Otherwise we have to expand
// it. We only expand quotes, and we don't want to do expensive expansions like tilde
// expansions. So we do our own "cleanliness" check; if we find a character not in our allowed
// set we know it's not a keyword, and if we never find a quote we don't have to expand! Note
// that this lowercase set could be shrunk to be just the characters that are in keywords.
let mut result = ParseKeyword::none;
let mut needs_expand = false;
let mut all_chars_valid = true;
for c in token.chars() {
if !is_keyword_char(c) {
all_chars_valid = false;
break;
}
// If we encounter a quote, we need expansion.
needs_expand = needs_expand || c == '"' || c == '\'' || c == '\\'
}
if all_chars_valid {
// Expand if necessary.
if !needs_expand {
result = ParseKeyword::from(token);
} else if let Some(unescaped) = unescape_string(token, UnescapeStringStyle::default()) {
result = ParseKeyword::from(&unescaped[..]);
}
}
result
}
#[test]
#[serial]
fn test_ast_parse() {
test_init();
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
let src = L!("echo");
let ast = Ast::parse(src, ParseTreeFlags::empty(), None);
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
assert!(!ast.any_error);
}
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
#[derive(Clone, Copy, Debug, Eq, PartialEq)]
pub enum Category {
branch,
leaf,
list,
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}
#[derive(Clone, Copy, Debug, Eq, PartialEq)]
pub enum Type {
token_base,
keyword_base,
redirection,
variable_assignment,
variable_assignment_list,
argument_or_redirection,
argument_or_redirection_list,
statement,
job_pipeline,
job_conjunction,
for_header,
while_header,
function_header,
begin_header,
block_statement,
if_clause,
elseif_clause,
elseif_clause_list,
else_clause,
if_statement,
case_item,
switch_statement,
decorated_statement,
not_statement,
job_continuation,
job_continuation_list,
job_conjunction_continuation,
andor_job,
andor_job_list,
freestanding_argument_list,
token_conjunction,
job_conjunction_continuation_list,
maybe_newlines,
token_pipe,
case_item_list,
argument,
argument_list,
job_list,
Port AST to Rust The translation is fairly direct though it adds some duplication, for example there are multiple "match" statements that mimic function overloading. Rust has no overloading, and we cannot have generic methods in the Node trait (due to a Rust limitation, the error is like "cannot be made into an object") so we include the type name in method names. Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor") that takes care of the AST traversal while the AST consumption remains in C++ for now. In future, "IndentVisitor" should absorb the entirety of "indent_visitor_t". This pattern requires that "fish_indent" be exposed includable header to the CXX bridge. Alternatively, we could define FFI wrappers for recursive AST traversal. Rust requires we separate the AST visitors for "mut" and "const" scenarios. Take this opportunity to concretize both visitors: The only client that requires mutable access is the populator. To match the structure of the C++ populator which makes heavy use of function overloading, we need to add a bunch of functions to the trait. Since there is no other mutable visit, this seems acceptable. The "const" visitors never use "will_visit_fields_of()" or "did_visit_fields_of()", so remove them (though this is debatable). Like in the C++ implementation, the AST nodes themselves are largely defined via macros. Union fields like "Statement" and "ArgumentOrRedirection" do currently not use macros but may in future. This commit also introduces a precedent for a type that is defined in one CXX bridge and used in another one - "ParseErrorList". To make this work we need to manually define "ExternType". There is one annoyance with CXX: functions that take explicit lifetime parameters require to be marked as unsafe. This makes little sense because functions that return `&Foo` with implicit lifetime can be misused the same way on the C++ side. One notable change is that we cannot directly port "find_block_open_keyword()" (which is used to compute an error) because it relies on the stack of visited nodes. We cannot modify a stack of node references while we do the "mut" walk. Happily, an idiomatic solution is easy: we can tell the AST visitor to backtrack to the parent node and create the error there. Since "node_t::accept_base" is no longer a template we don't need the "node_visitation_t" trampoline anymore. The added copying at the FFI boundary makes things slower (memcpy dominates the profile) but it's not unusable, which is good news: $ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'" Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish' Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms] Range (min … max): 193.2 ms … 205.1 ms 15 runs Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish' Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms] Range (min … max): 611.7 ms … 805.5 ms 10 runs Summary './fish.old -c 'source ../share/completions/git.fish'' ran 3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish'' Leftovers: - Enum variants are still snakecase; I didn't get around to changing this yet. - "ast_type_to_string()" still returns a snakecase name. This could be changed since it's not user visible.
2023-04-02 22:42:59 +08:00
}