fish-shell/src/parse_execution.cpp

1684 lines
68 KiB
C++
Raw Normal View History

// Provides the "linkage" between an ast and actual execution structures (job_t, etc.)
#include "config.h" // IWYU pragma: keep
#include "parse_execution.h"
2015-07-25 15:14:25 +00:00
#include <errno.h>
#include <stdarg.h>
#include <stdio.h>
#include <stdlib.h>
#include <termios.h>
#include <unistd.h>
#include <cwchar>
#include <deque>
#include <functional>
#include <memory>
#include <string>
#include <utility>
2015-07-25 15:14:25 +00:00
#include <vector>
#include "ast.h"
#include "builtin.h"
#include "builtins/function.h"
#include "common.h"
#include "complete.h"
2015-07-25 15:14:25 +00:00
#include "env.h"
#include "event.h"
#include "exec.h"
#include "expand.h"
#include "ffi.h"
2019-05-27 22:56:53 +00:00
#include "flog.h"
2015-07-25 15:14:25 +00:00
#include "function.h"
#include "io.h"
#include "job_group.rs.h"
#include "maybe.h"
#include "operation_context.h"
#include "parse_constants.h"
#include "parse_util.h"
#include "parser.h"
#include "path.h"
#include "proc.h"
#include "reader.h"
2023-02-14 21:54:18 +00:00
#include "timer.rs.h"
#include "tokenizer.h"
2023-03-28 15:59:51 +00:00
#include "trace.rs.h"
#include "wildcard.h"
#include "wutil.h"
/// These are the specific statement types that support redirections.
static constexpr bool type_is_redirectable_block(ast::type_t type) {
using t = ast::type_t;
return type == t::block_statement || type == t::if_statement || type == t::switch_statement;
}
static bool specific_statement_type_is_redirectable_block(const ast::node_t &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 14:42:59 +00:00
return type_is_redirectable_block(node.typ());
}
/// Get the name of a redirectable block, for profiling purposes.
static wcstring profiling_cmd_name_for_redirectable_block(const ast::node_t &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 14:42:59 +00:00
const parsed_source_ref_t &pstree) {
using namespace ast;
assert(specific_statement_type_is_redirectable_block(node));
2014-03-31 17:01:39 +00:00
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 14:42:59 +00:00
assert(node.try_source_range() && "No source range for block");
auto source_range = node.source_range();
size_t src_end = 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 14:42:59 +00:00
switch (node.typ()) {
case type_t::block_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 14:42:59 +00:00
auto block_header = node.as_block_statement().header().ptr();
switch (block_header->typ()) {
case type_t::for_header:
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 14:42:59 +00:00
src_end = block_header->as_for_header().semi_nl().source_range().start;
break;
case type_t::while_header:
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 14:42:59 +00:00
src_end =
block_header->as_while_header().condition().ptr()->source_range().end();
break;
case type_t::function_header:
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 14:42:59 +00:00
src_end = block_header->as_function_header().semi_nl().source_range().start;
break;
case type_t::begin_header:
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 14:42:59 +00:00
src_end =
block_header->as_begin_header().kw_begin().ptr()->source_range().end();
break;
default:
DIE("Unexpected block header type");
}
} break;
case type_t::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 14:42:59 +00:00
src_end =
node.as_if_statement().if_clause().condition().job().ptr()->source_range().end();
break;
case type_t::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 14:42:59 +00:00
src_end = node.as_switch_statement().semi_nl().source_range().start;
break;
default:
DIE("Not a redirectable block type");
break;
}
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 14:42:59 +00:00
assert(src_end >= source_range.start && "Invalid source end");
2014-03-31 17:01:39 +00:00
// Get the source for the block, and cut it at the next statement terminator.
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 14:42:59 +00:00
wcstring result = pstree.src().substr(source_range.start, src_end - source_range.start);
result.append(L"...");
return result;
}
/// Get a redirection from stderr to stdout (i.e. 2>&1).
2023-02-04 10:21:42 +00:00
static rust::Box<redirection_spec_t> get_stderr_merge() {
const wchar_t *stdout_fileno_str = L"1";
2023-02-04 10:21:42 +00:00
return new_redirection_spec(STDERR_FILENO, redirection_mode_t::fd, stdout_fileno_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 14:42:59 +00:00
parse_execution_context_t::parse_execution_context_t(rust::Box<parsed_source_ref_t> pstree,
const operation_context_t &ctx,
io_chain_t block_io)
: pstree(std::move(pstree)),
parser(ctx.parser.get()),
ctx(ctx),
block_io(std::move(block_io)) {}
// Utilities
wcstring parse_execution_context_t::get_source(const ast::node_t &node) const {
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 14:42:59 +00:00
return *node.source(pstree->src());
}
const ast::decorated_statement_t *
parse_execution_context_t::infinite_recursive_statement_in_job_list(const ast::job_list_t &jobs,
wcstring *out_func_name) const {
// This is a bit fragile. It is a test to see if we are inside of function call, but not inside
// a block in that function call. If, in the future, the rules for what block scopes are pushed
// on function invocation changes, then this check will break.
const block_t *current = parser->block_at_index(0), *parent = parser->block_at_index(1);
bool is_within_function_call =
2019-12-22 23:37:14 +00:00
(current && parent && current->type() == block_type_t::top && parent->is_function_call());
if (!is_within_function_call) {
return nullptr;
}
// Get the function name of the immediate block.
const wcstring &forbidden_function_name = parent->function_name;
// Get the first job in the job list.
const ast::job_conjunction_t *jc = jobs.at(0);
if (!jc) return nullptr;
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 14:42:59 +00:00
const ast::job_pipeline_t *job = &jc->job();
// Helper to return if a statement is infinitely recursive in this function.
auto statement_recurses =
[&](const ast::statement_t &stat) -> const ast::decorated_statement_t * {
// Ignore non-decorated statements like `if`, etc.
const ast::decorated_statement_t *dc =
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 14:42:59 +00:00
stat.contents().ptr()->try_as_decorated_statement()
? &stat.contents().ptr()->as_decorated_statement()
: nullptr;
if (!dc) return nullptr;
// Ignore statements with decorations like 'builtin' or 'command', since those
// are not infinite recursion. In particular that is what enables 'wrapper functions'.
if (dc->decoration() != statement_decoration_t::none) return nullptr;
// Check the command.
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 14:42:59 +00:00
wcstring cmd = *dc->command().source(pstree->src());
bool forbidden =
!cmd.empty() &&
expand_one(cmd, {expand_flag::skip_cmdsubst, expand_flag::skip_variables}, ctx) &&
cmd == forbidden_function_name;
return forbidden ? dc : nullptr;
};
const ast::decorated_statement_t *infinite_recursive_statement = nullptr;
// Check main 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 14:42:59 +00:00
infinite_recursive_statement = statement_recurses(jc->job().statement());
// Check piped remainder.
if (!infinite_recursive_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 14:42:59 +00:00
for (size_t i = 0; i < job->continuation().count(); i++) {
const ast::job_continuation_t &c = *job->continuation().at(i);
if (const auto *s = statement_recurses(c.statement())) {
infinite_recursive_statement = s;
2018-01-16 00:33:36 +00:00
break;
}
}
}
2018-01-16 00:33:36 +00:00
if (infinite_recursive_statement && out_func_name) {
*out_func_name = forbidden_function_name;
}
// may be null
return infinite_recursive_statement;
}
process_type_t parse_execution_context_t::process_type_for_command(
const ast::decorated_statement_t &statement, const wcstring &cmd) const {
enum process_type_t process_type = process_type_t::external;
// Determine the process type, which depends on the statement decoration (command, builtin,
// etc).
switch (statement.decoration()) {
case statement_decoration_t::exec:
process_type = process_type_t::exec;
break;
case statement_decoration_t::command:
process_type = process_type_t::external;
break;
case statement_decoration_t::builtin:
process_type = process_type_t::builtin;
break;
case statement_decoration_t::none:
if (function_exists(cmd, *parser)) {
process_type = process_type_t::function;
} else if (builtin_exists(cmd)) {
process_type = process_type_t::builtin;
} else {
process_type = process_type_t::external;
}
break;
}
return process_type;
}
maybe_t<end_execution_reason_t> parse_execution_context_t::check_end_execution() const {
// If one of our jobs ended with SIGINT, we stop execution.
// Likewise if fish itself got a SIGINT, or if something ran exit, etc.
if (cancel_signal || ctx.check_cancel() || fish_is_unwinding_for_exit()) {
return end_execution_reason_t::cancelled;
}
const auto &ld = parser->libdata();
if (ld.exit_current_script) {
return end_execution_reason_t::cancelled;
}
if (ld.returning) {
return end_execution_reason_t::control_flow;
}
if (ld.loop_status != loop_status_t::normals) {
return end_execution_reason_t::control_flow;
}
return none();
}
/// Return whether the job contains a single statement, of block type, with no redirections.
bool parse_execution_context_t::job_is_simple_block(const ast::job_pipeline_t &job) const {
using namespace ast;
// Must be no pipes.
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 14:42:59 +00:00
if (!job.continuation().empty()) {
return false;
}
// Helper to check if an argument_or_redirection_list_t has no redirections.
auto no_redirs = [](const argument_or_redirection_list_t &list) -> bool {
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 14:42:59 +00:00
for (size_t i = 0; i < list.count(); i++) {
const argument_or_redirection_t &val = *list.at(i);
if (val.is_redirection()) return false;
}
return true;
};
// Check if we're a block statement with redirections. We do it this obnoxious way to preserve
// type safety (in case we add more specific statement types).
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 14:42:59 +00:00
const auto ss = job.statement().contents().ptr();
switch (ss->typ()) {
case type_t::block_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 14:42:59 +00:00
return no_redirs(ss->as_block_statement().args_or_redirs());
case type_t::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 14:42:59 +00:00
return no_redirs(ss->as_switch_statement().args_or_redirs());
case type_t::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 14:42:59 +00:00
return no_redirs(ss->as_if_statement().args_or_redirs());
case type_t::not_statement:
case type_t::decorated_statement:
// not block statements
return false;
default:
assert(0 && "Unexpected child block type");
return false;
}
}
end_execution_reason_t parse_execution_context_t::run_if_statement(
const ast::if_statement_t &statement, const block_t *associated_block) {
using namespace ast;
using job_list_t = ast::job_list_t;
end_execution_reason_t result = end_execution_reason_t::ok;
// We have a sequence of if clauses, with a final else, resulting in a single job list that we
// execute.
const job_list_t *job_list_to_execute = nullptr;
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 14:42:59 +00:00
const if_clause_t *if_clause = &statement.if_clause();
// Index of the *next* elseif_clause to test.
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 14:42:59 +00:00
const elseif_clause_list_t &elseif_clauses = statement.elseif_clauses();
size_t next_elseif_idx = 0;
// We start with the 'if'.
trace_if_enabled(*parser, L"if");
for (;;) {
if (auto ret = check_end_execution()) {
result = *ret;
break;
}
// An if condition has a job and a "tail" of andor jobs, e.g. "foo ; and bar; or baz".
// Check the condition and the tail. We treat end_execution_reason_t::error here as failure,
// in accordance with historic behavior.
end_execution_reason_t cond_ret =
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 14:42:59 +00:00
run_job_conjunction(if_clause->condition(), associated_block);
if (cond_ret == end_execution_reason_t::ok) {
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 14:42:59 +00:00
cond_ret = run_job_list(if_clause->andor_tail(), associated_block);
}
const bool take_branch =
(cond_ret == end_execution_reason_t::ok) && parser->get_last_status() == EXIT_SUCCESS;
if (take_branch) {
// Condition 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 14:42:59 +00:00
job_list_to_execute = &if_clause->body();
break;
2018-01-14 10:02:02 +00:00
}
// See if we have an elseif.
const auto *elseif_clause = elseif_clauses.at(next_elseif_idx++);
if (elseif_clause) {
trace_if_enabled(*parser, L"else if");
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 14:42:59 +00:00
if_clause = &elseif_clause->if_clause();
} else {
break;
}
}
if (!job_list_to_execute) {
// our ifs and elseifs failed.
// Check our else body.
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 14:42:59 +00:00
if (statement.has_else_clause()) {
trace_if_enabled(*parser, L"else");
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 14:42:59 +00:00
job_list_to_execute = &statement.else_clause().body();
}
}
if (!job_list_to_execute) {
// 'if' condition failed, no else clause, return 0, we're done.
// No job list means no successful conditions, so return 0 (issue #1443).
parser->set_last_statuses(statuses_t::just(STATUS_CMD_OK));
} else {
// Execute the job list we got.
block_t *ib = parser->push_block(block_t::if_block());
run_job_list(*job_list_to_execute, ib);
if (auto ret = check_end_execution()) {
result = *ret;
}
parser->pop_block(ib);
}
trace_if_enabled(*parser, L"end if");
// It's possible there's a last-minute cancellation (issue #1297).
if (auto ret = check_end_execution()) {
result = *ret;
}
// Otherwise, take the exit status of the job list. Reversal of issue #1061.
return result;
}
end_execution_reason_t parse_execution_context_t::run_begin_statement(
const ast::job_list_t &contents) {
// Basic begin/end block. Push a scope block, run jobs, pop it
trace_if_enabled(*parser, L"begin");
2019-12-22 23:37:14 +00:00
block_t *sb = parser->push_block(block_t::scope_block(block_type_t::begin));
end_execution_reason_t ret = run_job_list(contents, sb);
parser->pop_block(sb);
trace_if_enabled(*parser, L"end begin");
return ret;
}
// Define a function.
end_execution_reason_t parse_execution_context_t::run_function_statement(
const ast::block_statement_t &statement, const ast::function_header_t &header) {
using namespace ast;
// Get arguments.
std::vector<wcstring> arguments;
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 14:42:59 +00:00
ast_args_list_t arg_nodes = get_argument_nodes(header.args());
arg_nodes.insert(arg_nodes.begin(), &header.first_arg());
end_execution_reason_t result =
this->expand_arguments_from_nodes(arg_nodes, &arguments, failglob);
if (result != end_execution_reason_t::ok) {
2016-10-31 03:26:10 +00:00
return result;
}
2023-03-28 15:59:51 +00:00
trace_if_enabled(*parser, L"function", arguments);
null_output_stream_t outs;
string_output_stream_t errs;
io_streams_t streams(outs, errs);
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 14:42:59 +00:00
int err_code = builtin_function(*parser, streams, arguments, *pstree, statement);
parser->libdata().status_count++;
parser->set_last_statuses(statuses_t::just(err_code));
2016-10-31 03:26:10 +00:00
const wcstring &errtext = errs.contents();
if (!errtext.empty()) {
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 14:42:59 +00:00
return this->report_error(err_code, *header.ptr(), L"%ls", errtext.c_str());
}
return result;
}
end_execution_reason_t parse_execution_context_t::run_block_statement(
const ast::block_statement_t &statement, const block_t *associated_block) {
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 14:42:59 +00:00
auto bh = statement.header().ptr();
const ast::job_list_t &contents = statement.jobs();
end_execution_reason_t ret = end_execution_reason_t::ok;
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 14:42:59 +00:00
if (const auto *fh = bh->try_as_for_header()) {
ret = run_for_statement(*fh, contents);
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 14:42:59 +00:00
} else if (const auto *wh = bh->try_as_while_header()) {
ret = run_while_statement(*wh, contents, associated_block);
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 14:42:59 +00:00
} else if (const auto *fh = bh->try_as_function_header()) {
ret = run_function_statement(statement, *fh);
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 14:42:59 +00:00
} else if (bh->try_as_begin_header()) {
2018-01-20 22:05:34 +00:00
ret = run_begin_statement(contents);
2018-01-14 09:42:58 +00:00
} else {
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 14:42:59 +00:00
FLOGF(error, L"Unexpected block header: %ls\n", bh->describe()->c_str());
2018-01-14 09:42:58 +00:00
PARSER_DIE();
}
return ret;
}
end_execution_reason_t parse_execution_context_t::run_for_statement(
const ast::for_header_t &header, const ast::job_list_t &block_contents) {
// Get the variable name: `for var_name in ...`. We expand the variable name. It better result
// in just one.
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 14:42:59 +00:00
wcstring for_var_name = *header.var_name().source(get_source());
if (!expand_one(for_var_name, expand_flags_t{}, ctx)) {
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 14:42:59 +00:00
return report_error(STATUS_EXPAND_ERROR, *header.var_name().ptr(),
FAILED_EXPANSION_VARIABLE_NAME_ERR_MSG, for_var_name.c_str());
}
if (!valid_var_name(for_var_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 14:42:59 +00:00
return report_error(STATUS_INVALID_ARGS, *header.var_name().ptr(), BUILTIN_ERR_VARNAME,
L"for", for_var_name.c_str());
}
// Get the contents to iterate over.
std::vector<wcstring> arguments;
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 14:42:59 +00:00
ast_args_list_t arg_nodes = get_argument_nodes(header.args());
end_execution_reason_t ret = this->expand_arguments_from_nodes(arg_nodes, &arguments, nullglob);
if (ret != end_execution_reason_t::ok) {
return ret;
}
auto var = parser->vars().get(for_var_name, ENV_DEFAULT);
if (env_var_t::flags_for(for_var_name.c_str()) & env_var_t::flag_read_only) {
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 14:42:59 +00:00
return report_error(STATUS_INVALID_ARGS, *header.var_name().ptr(),
2023-02-04 10:21:42 +00:00
_(L"%ls: %ls: cannot overwrite read-only variable"), L"for",
for_var_name.c_str());
}
auto &vars = parser->vars();
int retval;
retval = vars.set(for_var_name, ENV_LOCAL | ENV_USER,
var ? var->as_list() : std::vector<wcstring>{});
assert(retval == ENV_OK);
trace_if_enabled(*parser, L"for", arguments);
block_t *fb = parser->push_block(block_t::for_block());
// We fire the same event over and over again, just construct it once.
auto evt = new_event_variable_set(for_var_name);
// Now drive the for loop.
for (const wcstring &val : arguments) {
if (auto reason = check_end_execution()) {
ret = *reason;
break;
}
retval = vars.set(for_var_name, ENV_DEFAULT | ENV_USER, {val});
2017-12-21 21:53:18 +00:00
assert(retval == ENV_OK && "for loop variable should have been successfully set");
(void)retval;
event_fire(*parser, *evt);
auto &ld = parser->libdata();
ld.loop_status = loop_status_t::normals;
this->run_job_list(block_contents, fb);
if (check_end_execution() == end_execution_reason_t::control_flow) {
// Handle break or continue.
bool do_break = (ld.loop_status == loop_status_t::breaks);
ld.loop_status = loop_status_t::normals;
if (do_break) {
break;
}
}
}
parser->pop_block(fb);
trace_if_enabled(*parser, L"end for");
return ret;
}
end_execution_reason_t parse_execution_context_t::run_switch_statement(
const ast::switch_statement_t &statement) {
// Get the switch variable.
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 14:42:59 +00:00
const wcstring switch_value = get_source(*statement.argument().ptr());
// Expand it. We need to offset any errors by the position of the string.
completion_list_t switch_values_expanded;
2023-02-05 08:35:06 +00:00
auto errors = new_parse_error_list();
auto expand_ret =
2023-02-05 08:35:06 +00:00
expand_string(switch_value, &switch_values_expanded, expand_flags_t{}, ctx, &*errors);
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 14:42:59 +00:00
errors->offset_source_start(statement.argument().range().start);
2014-03-31 17:01:39 +00:00
switch (expand_ret.result) {
case expand_result_t::error:
2023-02-05 08:35:06 +00:00
return report_errors(expand_ret.status, *errors);
case expand_result_t::cancel:
return end_execution_reason_t::cancelled;
case expand_result_t::wildcard_no_match:
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 14:42:59 +00:00
return report_error(STATUS_UNMATCHED_WILDCARD, *statement.argument().ptr(),
WILDCARD_ERR_MSG, get_source(*statement.argument().ptr()).c_str());
2014-03-31 17:01:39 +00:00
case expand_result_t::ok:
if (switch_values_expanded.size() > 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 14:42:59 +00:00
return report_error(STATUS_INVALID_ARGS, *statement.argument().ptr(),
_(L"switch: Expected at most one argument, got %lu\n"),
switch_values_expanded.size());
}
break;
2016-10-31 03:26:10 +00:00
}
2014-03-31 17:01:39 +00:00
// If we expanded to nothing, match the empty string.
assert(switch_values_expanded.size() <= 1 && "Should have at most one expansion");
wcstring switch_value_expanded;
if (!switch_values_expanded.empty()) {
switch_value_expanded = std::move(switch_values_expanded.front().completion);
}
end_execution_reason_t result = end_execution_reason_t::ok;
2023-03-28 15:59:51 +00:00
trace_if_enabled(*parser, L"switch", {switch_value_expanded});
block_t *sb = parser->push_block(block_t::switch_block());
2016-10-31 03:26:10 +00:00
// Expand case statements.
const ast::case_item_t *matching_case_item = nullptr;
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 14:42:59 +00:00
for (size_t i = 0; i < statement.cases().count(); i++) {
const ast::case_item_t &case_item = *statement.cases().at(i);
if (auto ret = check_end_execution()) {
result = *ret;
2016-10-31 03:26:10 +00:00
break;
}
2016-10-31 03:26:10 +00:00
// Expand arguments. A case item list may have a wildcard that fails to expand to
// anything. We also report case errors, but don't stop execution; i.e. a case item that
// contains an unexpandable process will report and then fail to match.
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 14:42:59 +00:00
ast_args_list_t arg_nodes = get_argument_nodes(case_item.arguments());
std::vector<wcstring> case_args;
end_execution_reason_t case_result =
this->expand_arguments_from_nodes(arg_nodes, &case_args, failglob);
if (case_result == end_execution_reason_t::ok) {
2018-01-14 10:30:18 +00:00
for (const wcstring &arg : case_args) {
2016-10-31 03:26:10 +00:00
// Unescape wildcards so they can be expanded again.
wcstring unescaped_arg = parse_util_unescape_wildcards(arg);
bool match = wildcard_match(switch_value_expanded, unescaped_arg);
// If this matched, we're done.
if (match) {
matching_case_item = &case_item;
2016-10-31 03:26:10 +00:00
break;
}
}
}
2018-01-14 10:30:18 +00:00
if (matching_case_item) break;
2016-10-31 03:26:10 +00:00
}
2018-01-14 10:30:18 +00:00
if (matching_case_item) {
2016-10-31 03:26:10 +00:00
// Success, evaluate the job list.
assert(result == end_execution_reason_t::ok && "Expected success");
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 14:42:59 +00:00
result = this->run_job_list(matching_case_item->body(), sb);
}
2016-10-31 03:26:10 +00:00
parser->pop_block(sb);
trace_if_enabled(*parser, L"end switch");
return result;
}
end_execution_reason_t parse_execution_context_t::run_while_statement(
const ast::while_header_t &header, const ast::job_list_t &contents,
const block_t *associated_block) {
end_execution_reason_t ret = end_execution_reason_t::ok;
// "The exit status of the while loop shall be the exit status of the last compound-list-2
// executed, or zero if none was executed."
// Here are more detailed requirements:
// - If we execute the loop body zero times, or the loop body is empty, the status is success.
// - An empty loop body is treated as true, both in the loop condition and after loop exit.
// - The exit status of the last command is visible in the loop condition. (i.e. do not set the
// exit status to true BEFORE executing the loop condition).
// We achieve this by restoring the status if the loop condition fails, plus a special
// affordance for the first condition.
bool first_cond_check = true;
trace_if_enabled(*parser, L"while");
// Run while the condition is true.
for (;;) {
// Save off the exit status if it came from the loop body. We'll restore it if the condition
// is false.
auto cond_saved_status =
first_cond_check ? statuses_t::just(EXIT_SUCCESS) : parser->get_last_statuses();
first_cond_check = false;
// Check the condition.
end_execution_reason_t cond_ret =
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 14:42:59 +00:00
this->run_job_conjunction(header.condition(), associated_block);
if (cond_ret == end_execution_reason_t::ok) {
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 14:42:59 +00:00
cond_ret = run_job_list(header.andor_tail(), associated_block);
}
// If the loop condition failed to execute, then exit the loop without modifying the exit
// status. If the loop condition executed with a failure status, restore the status and then
// exit the loop.
if (cond_ret != end_execution_reason_t::ok) {
break;
} else if (parser->get_last_status() != EXIT_SUCCESS) {
parser->set_last_statuses(cond_saved_status);
break;
}
// Check cancellation.
if (auto reason = check_end_execution()) {
ret = *reason;
break;
}
// Push a while block and then check its cancellation reason.
auto &ld = parser->libdata();
ld.loop_status = loop_status_t::normals;
block_t *wb = parser->push_block(block_t::while_block());
2018-01-14 10:41:37 +00:00
this->run_job_list(contents, wb);
auto cancel_reason = this->check_end_execution();
parser->pop_block(wb);
if (cancel_reason == end_execution_reason_t::control_flow) {
// Handle break or continue.
bool do_break = (ld.loop_status == loop_status_t::breaks);
ld.loop_status = loop_status_t::normals;
if (do_break) {
break;
} else {
continue;
}
}
// no_exec means that fish was invoked with -n or --no-execute. If set, we allow the loop to
// not-execute once so its contents can be checked, and then break.
2019-05-12 22:04:18 +00:00
if (no_exec()) {
break;
}
}
trace_if_enabled(*parser, L"end while");
return ret;
}
// Reports an error. Always returns end_execution_reason_t::error.
end_execution_reason_t parse_execution_context_t::report_error(int status, const ast::node_t &node,
const wchar_t *fmt, ...) const {
auto r = node.source_range();
// Create an error.
2023-02-05 08:35:06 +00:00
auto error_list = new_parse_error_list();
parse_error_t error;
error.source_start = r.start;
error.source_length = r.length;
error.code = parse_error_code_t::syntax; // hackish
va_list va;
va_start(va, fmt);
2023-02-05 08:35:06 +00:00
error.text = std::make_unique<wcstring>(vformat_string(fmt, va));
va_end(va);
2023-02-05 08:35:06 +00:00
error_list->push_back(std::move(error));
return this->report_errors(status, *error_list);
}
end_execution_reason_t parse_execution_context_t::report_errors(
int status, const parse_error_list_t &error_list) const {
if (!ctx.check_cancel()) {
if (error_list.empty()) {
FLOG(error, L"Error reported but no error text found.");
}
2014-03-31 17:01:39 +00:00
// Get a backtrace.
wcstring backtrace_and_desc;
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 14:42:59 +00:00
parser->get_backtrace(pstree->src(), error_list, backtrace_and_desc);
2014-03-31 17:01:39 +00:00
// Print it.
if (!should_suppress_stderr_for_tests()) {
std::fwprintf(stderr, L"%ls", backtrace_and_desc.c_str());
}
// Mark status.
parser->set_last_statuses(statuses_t::just(status));
}
return end_execution_reason_t::error;
}
// static
parse_execution_context_t::ast_args_list_t parse_execution_context_t::get_argument_nodes(
const ast::argument_list_t &args) {
ast_args_list_t 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 14:42:59 +00:00
for (size_t i = 0; i < args.count(); i++) {
const ast::argument_t &arg = *args.at(i);
result.push_back(&arg);
}
return result;
}
// static
parse_execution_context_t::ast_args_list_t parse_execution_context_t::get_argument_nodes(
const ast::argument_or_redirection_list_t &args) {
ast_args_list_t 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 14:42:59 +00:00
for (size_t i = 0; i < args.count(); i++) {
const ast::argument_or_redirection_t &v = *args.at(i);
if (v.is_argument()) result.push_back(&v.argument());
}
return result;
}
/// Handle the case of command not found.
end_execution_reason_t parse_execution_context_t::handle_command_not_found(
const wcstring &cmd_str, const ast::decorated_statement_t &statement, int err_code) {
// We couldn't find the specified command. This is a non-fatal error. We want to set the exit
// status to 127, which is the standard number used by other shells like bash and zsh.
const wchar_t *const cmd = cmd_str.c_str();
if (err_code != ENOENT) {
// TODO: We currently handle all errors here the same,
// but this mainly applies to EACCES. We could also feasibly get:
// ELOOP
// ENAMETOOLONG
if (err_code == ENOTDIR) {
// If the original command did not include a "/", assume we found it via $PATH.
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 14:42:59 +00:00
auto src = get_source(*statement.command().ptr());
if (src.find(L"/") == wcstring::npos) {
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 14:42:59 +00:00
return this->report_error(STATUS_NOT_EXECUTABLE, *statement.command().ptr(),
2023-02-04 10:21:42 +00:00
_(L"Unknown command. A component of '%ls' is not a "
L"directory. Check your $PATH."),
cmd);
} else {
return this->report_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 14:42:59 +00:00
STATUS_NOT_EXECUTABLE, *statement.command().ptr(),
2023-02-04 10:21:42 +00:00
_(L"Unknown command. A component of '%ls' is not a directory."), cmd);
}
}
return this->report_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 14:42:59 +00:00
STATUS_NOT_EXECUTABLE, *statement.command().ptr(),
2023-02-04 10:21:42 +00:00
_(L"Unknown command. '%ls' exists but is not an executable file."), cmd);
}
// Handle unrecognized commands with standard command not found handler that can make better
// error messages.
std::vector<wcstring> event_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 14:42:59 +00:00
ast_args_list_t args = get_argument_nodes(statement.args_or_redirs());
end_execution_reason_t arg_result =
this->expand_arguments_from_nodes(args, &event_args, failglob);
if (arg_result != end_execution_reason_t::ok) {
return arg_result;
}
event_args.insert(event_args.begin(), cmd_str);
}
wcstring buffer;
wcstring error;
// Redirect to stderr
auto io = io_chain_t{};
2023-02-04 10:21:42 +00:00
auto list = new_redirection_spec_list();
list->push_back(new_redirection_spec(STDOUT_FILENO, redirection_mode_t::fd, L"2"));
io.append_from_specs(*list, L"");
if (function_exists(L"fish_command_not_found", *parser)) {
buffer = L"fish_command_not_found";
for (const wcstring &arg : event_args) {
buffer.push_back(L' ');
buffer.append(escape_string(arg));
}
auto prev_statuses = parser->get_last_statuses();
auto event = new_event_generic(L"fish_command_not_found");
block_t *b = parser->push_block(block_t::event_block(&*event));
parser->eval(buffer, io);
parser->pop_block(b);
parser->set_last_statuses(std::move(prev_statuses));
} else {
// If we have no handler, just print it as a normal error.
error = _(L"Unknown command:");
if (!event_args.empty()) {
error.push_back(L' ');
error.append(escape_string(event_args[0]));
}
}
if (!cmd_str.empty() && cmd_str.at(0) == L'{') {
error.append(ERROR_NO_BRACE_GROUPING);
}
// Here we want to report an error (so it shows a backtrace).
// If the handler printed text, that's already shown, so error will be empty.
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 14:42:59 +00:00
return this->report_error(STATUS_CMD_UNKNOWN, *statement.command().ptr(), error.c_str());
}
end_execution_reason_t parse_execution_context_t::expand_command(
const ast::decorated_statement_t &statement, wcstring *out_cmd,
std::vector<wcstring> *out_args) const {
// Here we're expanding a command, for example $HOME/bin/stuff or $randomthing. The first
// completion becomes the command itself, everything after becomes arguments. Command
// substitutions are not supported.
2023-02-05 08:35:06 +00:00
auto errors = new_parse_error_list();
// Get the unexpanded command string. We expect to always get it here.
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 14:42:59 +00:00
wcstring unexp_cmd = get_source(*statement.command().ptr());
size_t pos_of_command_token = statement.command().range().start;
// Expand the string to produce completions, and report errors.
expand_result_t expand_err =
2023-02-05 08:35:06 +00:00
expand_to_command_and_args(unexp_cmd, ctx, out_cmd, out_args, &*errors);
if (expand_err == expand_result_t::error) {
// Issue #5812 - the expansions were done on the command token,
// excluding prefixes such as " " or "if ".
// This means that the error positions are relative to the beginning
// of the token; we need to make them relative to the original source.
2023-02-05 08:35:06 +00:00
errors->offset_source_start(pos_of_command_token);
return report_errors(STATUS_ILLEGAL_CMD, *errors);
} else if (expand_err == expand_result_t::wildcard_no_match) {
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 14:42:59 +00:00
return report_error(STATUS_UNMATCHED_WILDCARD, *statement.ptr(), WILDCARD_ERR_MSG,
get_source(*statement.ptr()).c_str());
}
assert(expand_err == expand_result_t::ok);
// Complain if the resulting expansion was empty, or expanded to an empty string.
// For no-exec it's okay, as we can't really perform the expansion.
if (out_cmd->empty() && !no_exec()) {
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 14:42:59 +00:00
return this->report_error(STATUS_ILLEGAL_CMD, *statement.command().ptr(),
_(L"The expanded command was empty."));
}
return end_execution_reason_t::ok;
}
/// Creates a 'normal' (non-block) process.
end_execution_reason_t parse_execution_context_t::populate_plain_process(
process_t *proc, const ast::decorated_statement_t &statement) {
assert(proc != nullptr);
// We may decide that a command should be an implicit cd.
bool use_implicit_cd = false;
// Get the command and any arguments due to expanding the command.
wcstring cmd;
std::vector<wcstring> args_from_cmd_expansion;
auto ret = expand_command(statement, &cmd, &args_from_cmd_expansion);
if (ret != end_execution_reason_t::ok) {
return ret;
}
// For no-exec, having an empty command is okay. We can't do anything more with it tho.
if (no_exec()) return end_execution_reason_t::ok;
assert(!cmd.empty() && "expand_command should not produce an empty command");
// Determine the process type.
enum process_type_t process_type = process_type_for_command(statement, cmd);
get_path_result_t external_cmd{};
if (process_type == process_type_t::external || process_type == process_type_t::exec) {
// Determine the actual command. This may be an implicit cd.
external_cmd = path_try_get_path(cmd, parser->vars());
bool has_command = external_cmd.err == 0;
// If the specified command does not exist, and is undecorated, try using an implicit cd.
if (!has_command && statement.decoration() == statement_decoration_t::none) {
// Implicit cd requires an empty argument and redirection 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 14:42:59 +00:00
if (statement.args_or_redirs().empty()) {
// Ok, no arguments or redirections; check to see if the command is a directory.
use_implicit_cd =
path_as_implicit_cd(cmd, parser->vars().get_pwd_slash(), parser->vars())
.has_value();
}
}
if (!has_command && !use_implicit_cd) {
// No command. If we're --no-execute return okay - it might be a function.
if (no_exec()) return end_execution_reason_t::ok;
return this->handle_command_not_found(
external_cmd.path.empty() ? cmd : external_cmd.path, statement, external_cmd.err);
}
}
// Produce the full argument list and the set of IO redirections.
std::vector<wcstring> cmd_args;
2023-02-04 10:21:42 +00:00
auto redirections = new_redirection_spec_list();
if (use_implicit_cd) {
// Implicit cd is simple.
cmd_args = {L"cd", cmd};
external_cmd = get_path_result_t{};
// If we have defined a wrapper around cd, use it, otherwise use the cd builtin.
process_type =
function_exists(L"cd", *parser) ? process_type_t::function : process_type_t::builtin;
} else {
// Not implicit cd.
const globspec_t glob_behavior =
(cmd == L"set" || cmd == L"count" || cmd == L"path") ? nullglob : failglob;
// Form the list of arguments. The command is the first argument, followed by any arguments
// from expanding the command, followed by the argument nodes themselves. E.g. if the
// command is '$gco foo' and $gco is git checkout.
cmd_args.push_back(cmd);
vec_append(cmd_args, std::move(args_from_cmd_expansion));
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 14:42:59 +00:00
ast_args_list_t arg_nodes = get_argument_nodes(statement.args_or_redirs());
end_execution_reason_t arg_result =
this->expand_arguments_from_nodes(arg_nodes, &cmd_args, glob_behavior);
if (arg_result != end_execution_reason_t::ok) {
return arg_result;
}
// The set of IO redirections that we construct for the process.
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 14:42:59 +00:00
auto reason = this->determine_redirections(statement.args_or_redirs(), &*redirections);
if (reason != end_execution_reason_t::ok) {
return reason;
}
}
// Populate the process.
proc->type = process_type;
proc->set_argv(std::move(cmd_args));
proc->set_redirection_specs(std::move(redirections));
proc->actual_cmd = std::move(external_cmd.path);
return end_execution_reason_t::ok;
}
// Determine the list of arguments, expanding stuff. Reports any errors caused by expansion. If we
// have a wildcard that could not be expanded, report the error and continue.
end_execution_reason_t parse_execution_context_t::expand_arguments_from_nodes(
const ast_args_list_t &argument_nodes, std::vector<wcstring> *out_arguments,
globspec_t glob_behavior) {
// Get all argument nodes underneath the statement. We guess we'll have that many arguments (but
// may have more or fewer, if there are wildcards involved).
out_arguments->reserve(out_arguments->size() + argument_nodes.size());
completion_list_t arg_expanded;
for (const ast::argument_t *arg_node : argument_nodes) {
// Expect all arguments to have source.
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 14:42:59 +00:00
assert(arg_node->ptr()->has_source() && "Argument should have source");
// Expand this string.
2023-02-05 08:35:06 +00:00
auto errors = new_parse_error_list();
arg_expanded.clear();
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 14:42:59 +00:00
auto expand_ret = expand_string(get_source(*arg_node->ptr()), &arg_expanded,
expand_flags_t{}, ctx, &*errors);
errors->offset_source_start(arg_node->range().start);
switch (expand_ret.result) {
case expand_result_t::error: {
2023-02-05 08:35:06 +00:00
return this->report_errors(expand_ret.status, *errors);
}
case expand_result_t::cancel: {
return end_execution_reason_t::cancelled;
}
case expand_result_t::wildcard_no_match: {
if (glob_behavior == failglob) {
// For no_exec, ignore the error - this might work at runtime.
if (no_exec()) return end_execution_reason_t::ok;
// Report the unmatched wildcard error and stop processing.
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 14:42:59 +00:00
return report_error(STATUS_UNMATCHED_WILDCARD, *arg_node->ptr(),
WILDCARD_ERR_MSG, get_source(*arg_node->ptr()).c_str());
}
break;
}
case expand_result_t::ok: {
break;
}
default: {
DIE("unexpected expand_string() return value");
}
}
// Now copy over any expanded arguments. Use std::move() to avoid extra allocations; this
// is called very frequently.
out_arguments->reserve(out_arguments->size() + arg_expanded.size());
for (completion_t &new_arg : arg_expanded) {
out_arguments->push_back(std::move(new_arg.completion));
}
}
// We may have received a cancellation during this expansion.
if (auto ret = check_end_execution()) {
return *ret;
}
return end_execution_reason_t::ok;
}
end_execution_reason_t parse_execution_context_t::determine_redirections(
const ast::argument_or_redirection_list_t &list, redirection_spec_list_t *out_redirections) {
// Get all redirection nodes underneath the 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 14:42:59 +00:00
for (size_t i = 0; i < list.count(); i++) {
const ast::argument_or_redirection_t &arg_or_redir = *list.at(i);
if (!arg_or_redir.is_redirection()) continue;
const ast::redirection_t &redir_node = arg_or_redir.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 14:42:59 +00:00
auto oper = pipe_or_redir_from_string(get_source(*redir_node.oper().ptr()).c_str());
if (!oper || !oper->is_valid()) {
// TODO: figure out if this can ever happen. If so, improve this error message.
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 14:42:59 +00:00
return report_error(STATUS_INVALID_ARGS, *redir_node.ptr(),
_(L"Invalid redirection: %ls"),
get_source(*redir_node.ptr()).c_str());
}
// PCA: I can't justify this skip_variables flag. It was like this when I got here.
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 14:42:59 +00:00
wcstring target = get_source(*redir_node.target().ptr());
bool target_expanded =
expand_one(target, no_exec() ? expand_flag::skip_variables : expand_flags_t{}, ctx);
if (!target_expanded || target.empty()) {
// TODO: Improve this error message.
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 14:42:59 +00:00
return report_error(STATUS_INVALID_ARGS, *redir_node.ptr(),
_(L"Invalid redirection target: %ls"), target.c_str());
}
// Make a redirection spec from the redirect token.
assert(oper && oper->is_valid() && "expected to have a valid redirection");
2023-02-04 10:21:42 +00:00
auto spec = new_redirection_spec(oper->fd, oper->mode, target.c_str());
// Validate this spec.
2023-02-04 10:21:42 +00:00
if (spec->mode() == redirection_mode_t::fd && !spec->is_close() &&
!spec->get_target_as_fd()) {
const wchar_t *fmt =
_(L"Requested redirection to '%ls', which is not a valid file descriptor");
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 14:42:59 +00:00
return report_error(STATUS_INVALID_ARGS, *redir_node.ptr(), fmt,
spec->target()->c_str());
}
out_redirections->push_back(std::move(spec));
if (oper->stderr_merge) {
// This was a redirect like &> which also modifies stderr.
// Also redirect stderr to stdout.
out_redirections->push_back(get_stderr_merge());
}
}
return end_execution_reason_t::ok;
}
end_execution_reason_t parse_execution_context_t::populate_not_process(
job_t *job, process_t *proc, const ast::not_statement_t &not_statement) {
auto &flags = job->mut_flags();
flags.negate = !flags.negate;
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 14:42:59 +00:00
return this->populate_job_process(job, proc, not_statement.contents(),
not_statement.variables());
}
template <typename Type>
end_execution_reason_t parse_execution_context_t::populate_block_process(
process_t *proc, const ast::statement_t &statement, const Type &specific_statement) {
using namespace ast;
// We handle block statements by creating process_type_t::block_node, that will bounce back to
// us when it's time to execute them.
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 14:42:59 +00:00
static_assert(std::is_same<Type, block_statement_t>::value ||
std::is_same<Type, if_statement_t>::value ||
std::is_same<Type, switch_statement_t>::value,
"Invalid block process");
// Get the argument or redirections list.
// TODO: args_or_redirs should be available without resolving the statement type.
const argument_or_redirection_list_t *args_or_redirs = nullptr;
// Upcast to permit dropping the 'template' 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 14:42:59 +00:00
const auto ss = specific_statement.ptr();
switch (ss->typ()) {
case type_t::block_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 14:42:59 +00:00
args_or_redirs = &ss->as_block_statement().args_or_redirs();
break;
case type_t::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 14:42:59 +00:00
args_or_redirs = &ss->as_if_statement().args_or_redirs();
break;
case type_t::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 14:42:59 +00:00
args_or_redirs = &ss->as_switch_statement().args_or_redirs();
break;
default:
DIE("Unexpected block node type");
}
assert(args_or_redirs && "Should have args_or_redirs");
2023-02-04 10:21:42 +00:00
auto redirections = new_redirection_spec_list();
auto reason = this->determine_redirections(*args_or_redirs, &*redirections);
if (reason == end_execution_reason_t::ok) {
proc->type = process_type_t::block_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 14:42:59 +00:00
proc->block_node_source = pstree->clone();
proc->internal_block_node = &statement;
proc->set_redirection_specs(std::move(redirections));
}
return reason;
}
end_execution_reason_t parse_execution_context_t::apply_variable_assignments(
process_t *proc, const ast::variable_assignment_list_t &variable_assignment_list,
const block_t **block) {
if (variable_assignment_list.empty()) return end_execution_reason_t::ok;
*block = parser->push_block(block_t::variable_assignment_block());
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 14:42:59 +00:00
for (size_t i = 0; i < variable_assignment_list.count(); i++) {
const ast::variable_assignment_t &variable_assignment = *variable_assignment_list.at(i);
const wcstring &source = get_source(*variable_assignment.ptr());
auto equals_pos = variable_assignment_equals_pos(source);
2023-02-05 08:35:06 +00:00
assert(equals_pos);
const wcstring variable_name = source.substr(0, *equals_pos);
const wcstring expression = source.substr(*equals_pos + 1);
completion_list_t expression_expanded;
2023-02-05 08:35:06 +00:00
auto errors = new_parse_error_list();
// TODO this is mostly copied from expand_arguments_from_nodes, maybe extract to function
auto expand_ret =
2023-02-05 08:35:06 +00:00
expand_string(expression, &expression_expanded, expand_flags_t{}, ctx, &*errors);
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 14:42:59 +00:00
errors->offset_source_start(variable_assignment.range().start + *equals_pos + 1);
switch (expand_ret.result) {
case expand_result_t::error:
2023-02-05 08:35:06 +00:00
return this->report_errors(expand_ret.status, *errors);
case expand_result_t::cancel:
return end_execution_reason_t::cancelled;
case expand_result_t::wildcard_no_match: // nullglob (equivalent to set)
case expand_result_t::ok:
break;
default: {
DIE("unexpected expand_string() return value");
}
}
std::vector<wcstring> vals;
for (auto &completion : expression_expanded) {
vals.emplace_back(std::move(completion.completion));
}
if (proc) proc->variable_assignments.push_back({variable_name, vals});
parser->set_var_and_fire(variable_name, ENV_LOCAL | ENV_EXPORT, std::move(vals));
}
return end_execution_reason_t::ok;
}
end_execution_reason_t parse_execution_context_t::populate_job_process(
job_t *job, process_t *proc, const ast::statement_t &statement,
const ast::variable_assignment_list_t &variable_assignments) {
using namespace ast;
// Get the "specific statement" which is boolean / block / if / switch / decorated.
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 14:42:59 +00:00
const auto specific_statement = statement.contents().ptr();
const block_t *block = nullptr;
end_execution_reason_t result =
this->apply_variable_assignments(proc, variable_assignments, &block);
cleanup_t scope([&]() {
if (block) parser->pop_block(block);
});
if (result != end_execution_reason_t::ok) return 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 14:42:59 +00:00
switch (specific_statement->typ()) {
case type_t::not_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 14:42:59 +00:00
result = this->populate_not_process(job, proc, specific_statement->as_not_statement());
break;
}
case type_t::block_statement:
result = this->populate_block_process(proc, 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 14:42:59 +00:00
specific_statement->as_block_statement());
break;
case type_t::if_statement:
result = this->populate_block_process(proc, 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 14:42:59 +00:00
specific_statement->as_if_statement());
break;
case type_t::switch_statement:
result = this->populate_block_process(proc, 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 14:42:59 +00:00
specific_statement->as_switch_statement());
break;
case type_t::decorated_statement: {
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 14:42:59 +00:00
this->populate_plain_process(proc, specific_statement->as_decorated_statement());
break;
}
default: {
FLOGF(error, L"'%ls' not handled by new parser yet.",
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 14:42:59 +00:00
specific_statement->describe()->c_str());
PARSER_DIE();
break;
}
}
return result;
}
end_execution_reason_t parse_execution_context_t::populate_job_from_job_node(
job_t *j, const ast::job_pipeline_t &job_node, const block_t *associated_block) {
UNUSED(associated_block);
// We are going to construct process_t structures for every statement in the job.
// Create processes. Each one may fail.
process_list_t processes;
processes.emplace_back(new process_t());
end_execution_reason_t result = this->populate_job_process(
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 14:42:59 +00:00
j, processes.back().get(), job_node.statement(), job_node.variables());
// Construct process_ts for job continuations (pipelines).
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 14:42:59 +00:00
for (size_t i = 0; i < job_node.continuation().count(); i++) {
const ast::job_continuation_t &jc = *job_node.continuation().at(i);
if (result != end_execution_reason_t::ok) {
break;
}
// Handle the pipe, whose fd may not be the obvious stdout.
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 14:42:59 +00:00
auto parsed_pipe = pipe_or_redir_from_string(get_source(*jc.pipe().ptr()).c_str());
assert(parsed_pipe && parsed_pipe->is_pipe && "Failed to parse valid pipe");
if (!parsed_pipe->is_valid()) {
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 14:42:59 +00:00
result = report_error(STATUS_INVALID_ARGS, *jc.pipe().ptr(), ILLEGAL_FD_ERR_MSG,
get_source(*jc.pipe().ptr()).c_str());
break;
}
processes.back()->pipe_write_fd = parsed_pipe->fd;
if (parsed_pipe->stderr_merge) {
// This was a pipe like &| which redirects both stdout and stderr.
// Also redirect stderr to stdout.
2023-02-04 10:21:42 +00:00
auto specs = processes.back()->redirection_specs().clone();
specs->push_back(get_stderr_merge());
processes.back()->set_redirection_specs(std::move(specs));
}
// Store the new process (and maybe with an error).
processes.emplace_back(new process_t());
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 14:42:59 +00:00
result =
this->populate_job_process(j, processes.back().get(), jc.statement(), jc.variables());
}
// Inform our processes of who is first and last
processes.front()->is_first_in_job = true;
processes.back()->is_last_in_job = true;
// Return what happened.
if (result == end_execution_reason_t::ok) {
// Link up the processes.
2016-10-23 20:58:12 +00:00
assert(!processes.empty()); //!OCLINT(multiple unary operator)
j->processes = std::move(processes);
}
return result;
}
static bool remove_job(parser_t &parser, const job_t *job) {
for (auto j = parser.jobs().begin(); j != parser.jobs().end(); ++j) {
if (j->get() == job) {
parser.jobs().erase(j);
return true;
}
}
return false;
}
/// Decide if a job node should be 'time'd.
/// For historical reasons the 'not' and 'time' prefix are "inside out". That is, it's
/// 'not time cmd'. Note that a time appearing anywhere in the pipeline affects the whole job.
/// `sleep 1 | not time true` will time the whole job!
static bool job_node_wants_timing(const ast::job_pipeline_t &job_node) {
// Does our job have the job-level time prefix?
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 14:42:59 +00:00
if (job_node.has_time()) return true;
// Helper to return true if a node is 'not time ...' or 'not not time...' or...
auto is_timed_not_statement = [](const ast::statement_t &stat) {
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 14:42:59 +00:00
const auto *ns = stat.contents().ptr()->try_as_not_statement()
? &stat.contents().ptr()->as_not_statement()
: nullptr;
while (ns) {
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 14:42:59 +00:00
if (ns->has_time()) return true;
ns = ns->contents().ptr()->try_as_not_statement()
? &ns->contents().ptr()->as_not_statement()
: nullptr;
}
return false;
};
// Do we have a 'not time ...' anywhere in our pipeline?
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 14:42:59 +00:00
if (is_timed_not_statement(job_node.statement())) return true;
for (size_t i = 0; i < job_node.continuation().count(); i++) {
const ast::job_continuation_t &jc = *job_node.continuation().at(i);
if (is_timed_not_statement(jc.statement())) return true;
}
return false;
}
end_execution_reason_t parse_execution_context_t::run_1_job(const ast::job_pipeline_t &job_node,
const block_t *associated_block) {
if (auto ret = check_end_execution()) {
return *ret;
}
// We definitely do not want to execute anything if we're told we're --no-execute!
if (no_exec()) return end_execution_reason_t::ok;
// Get terminal modes.
struct termios tmodes = {};
if (parser->is_interactive() && tcgetattr(STDIN_FILENO, &tmodes)) {
// Need real error handling here.
wperror(L"tcgetattr");
parser->set_last_statuses(statuses_t::just(STATUS_CMD_ERROR));
return end_execution_reason_t::error;
}
// Increment the eval_level for the duration of this command.
scoped_push<int> saved_eval_level(&parser->eval_level, parser->eval_level + 1);
2014-03-31 17:01:39 +00:00
// Save the node index.
scoped_push<const ast::job_pipeline_t *> saved_node(&executing_job_node, &job_node);
// Profiling support.
profile_item_t *profile_item = this->parser->create_profile_item();
const auto start_time = profile_item ? profile_item_t::now() : 0;
// When we encounter a block construct (e.g. while loop) in the general case, we create a "block
// process" containing its node. This allows us to handle block-level redirections.
// However, if there are no redirections, then we can just jump into the block directly, which
// is significantly faster.
if (job_is_simple_block(job_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 14:42:59 +00:00
bool do_time = job_node.has_time();
// If no-exec has been given, there is nothing to time.
2023-02-14 21:54:18 +00:00
auto timer = push_timer(do_time && !no_exec());
const block_t *block = nullptr;
end_execution_reason_t 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 14:42:59 +00:00
this->apply_variable_assignments(nullptr, job_node.variables(), &block);
cleanup_t scope([&]() {
if (block) parser->pop_block(block);
});
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 14:42:59 +00:00
const auto specific_statement = job_node.statement().contents().ptr();
assert(specific_statement_type_is_redirectable_block(*specific_statement));
if (result == end_execution_reason_t::ok) {
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 14:42:59 +00:00
switch (specific_statement->typ()) {
case ast::type_t::block_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 14:42:59 +00:00
result = this->run_block_statement(specific_statement->as_block_statement(),
associated_block);
break;
}
case ast::type_t::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 14:42:59 +00:00
result = this->run_if_statement(specific_statement->as_if_statement(),
associated_block);
break;
}
case ast::type_t::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 14:42:59 +00:00
result = this->run_switch_statement(specific_statement->as_switch_statement());
break;
}
default: {
// Other types should be impossible due to the
// specific_statement_type_is_redirectable_block check.
PARSER_DIE();
break;
}
}
}
2014-03-31 17:01:39 +00:00
if (profile_item != nullptr) {
profile_item->duration = profile_item_t::now() - start_time;
profile_item->level = parser->eval_level;
profile_item->cmd =
profiling_cmd_name_for_redirectable_block(*specific_statement, *this->pstree);
profile_item->skipped = false;
}
2014-03-31 17:01:39 +00:00
return result;
}
const auto &ld = parser->libdata();
job_t::properties_t props{};
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 14:42:59 +00:00
props.initial_background = job_node.has_bg();
props.skip_notification =
ld.is_subshell || parser->is_block() || ld.is_event || !parser->is_interactive();
props.from_event_handler = ld.is_event;
props.wants_timing = job_node_wants_timing(job_node);
// It's an error to have 'time' in a background job.
if (props.wants_timing && props.initial_background) {
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 14:42:59 +00:00
return this->report_error(STATUS_INVALID_ARGS, *job_node.ptr(), ERROR_TIME_BACKGROUND);
}
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 14:42:59 +00:00
shared_ptr<job_t> job = std::make_shared<job_t>(props, get_source(*job_node.ptr()));
// We are about to populate a job. One possible argument to the job is a command substitution
// which may be interested in the job that's populating it, via '--on-job-exit caller'. Record
// the job ID here.
scoped_push<internal_job_id_t> caller_id(&parser->libdata().caller_id, job->internal_job_id);
// Populate the job. This may fail for reasons like command_not_found. If this fails, an error
// will have been printed.
end_execution_reason_t pop_result =
2018-01-16 00:00:19 +00:00
this->populate_job_from_job_node(job.get(), job_node, associated_block);
caller_id.restore();
// Clean up the job on failure or cancellation.
if (pop_result == end_execution_reason_t::ok) {
this->setup_group(job.get());
assert(job->group && "Should not have a null group");
// Give the job to the parser - it will clean it up.
parser->job_add(job);
// Actually execute the job.
if (!exec_job(*parser, job, block_io)) {
// No process in the job successfully launched.
// Ensure statuses are set (#7540).
if (auto statuses = job->get_statuses()) {
parser->set_last_statuses(statuses.value());
parser->libdata().status_count++;
}
remove_job(*this->parser, job.get());
}
2021-10-27 21:16:32 +00:00
// Update universal variables on external commands.
// We only incorporate external changes if we had an external proc, for hysterical raisins.
parser->sync_uvars_and_fire(job->has_external_proc() /* always */);
// If the job got a SIGINT or SIGQUIT, then we're going to start unwinding.
if (!cancel_signal) cancel_signal = job->group->get_cancel_signal();
}
if (profile_item != nullptr) {
profile_item->duration = profile_item_t::now() - start_time;
profile_item->level = parser->eval_level;
profile_item->cmd = job ? job->command() : wcstring();
profile_item->skipped = (pop_result != end_execution_reason_t::ok);
}
job_reap(*parser, false); // clean up jobs
return pop_result;
}
end_execution_reason_t parse_execution_context_t::run_job_conjunction(
const ast::job_conjunction_t &job_expr, const block_t *associated_block) {
if (auto reason = check_end_execution()) {
return *reason;
}
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 14:42:59 +00:00
end_execution_reason_t result = run_1_job(job_expr.job(), associated_block);
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 14:42:59 +00:00
for (size_t i = 0; i < job_expr.continuations().count(); i++) {
const ast::job_conjunction_continuation_t &jc = *job_expr.continuations().at(i);
if (result != end_execution_reason_t::ok) {
return result;
}
if (auto reason = check_end_execution()) {
return *reason;
}
// Check the conjunction type.
bool skip = false;
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 14:42:59 +00:00
switch (jc.conjunction().token_type()) {
case parse_token_type_t::andand:
// AND. Skip if the last job failed.
skip = parser->get_last_status() != 0;
break;
case parse_token_type_t::oror:
// OR. Skip if the last job succeeded.
skip = parser->get_last_status() == 0;
break;
default:
DIE("Unexpected job conjunction type");
}
if (!skip) {
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 14:42:59 +00:00
result = run_1_job(jc.job(), associated_block);
}
}
return result;
}
end_execution_reason_t parse_execution_context_t::test_and_run_1_job_conjunction(
const ast::job_conjunction_t &jc, const block_t *associated_block) {
// Test this job conjunction if it has an 'and' or 'or' decorator.
// If it passes, then run it.
if (auto reason = check_end_execution()) {
return *reason;
}
// Maybe skip the job if it has a leading and/or.
bool skip = false;
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 14:42:59 +00:00
if (jc.has_decorator()) {
2023-04-19 16:42:26 +00:00
switch (jc.decorator().keyword()) {
case parse_keyword_t::kw_and:
// AND. Skip if the last job failed.
skip = parser->get_last_status() != 0;
break;
case parse_keyword_t::kw_or:
// OR. Skip if the last job succeeded.
skip = parser->get_last_status() == 0;
break;
default:
DIE("Unexpected keyword");
}
}
// Skipping is treated as success.
if (skip) {
return end_execution_reason_t::ok;
} else {
return this->run_job_conjunction(jc, associated_block);
}
}
end_execution_reason_t parse_execution_context_t::run_job_list(const ast::job_list_t &job_list_node,
const block_t *associated_block) {
auto result = end_execution_reason_t::ok;
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 14:42:59 +00:00
for (size_t i = 0; i < job_list_node.count(); i++) {
const ast::job_conjunction_t *jc = job_list_node.at(i);
result = test_and_run_1_job_conjunction(*jc, associated_block);
}
// Returns the result of the last job executed or skipped.
return result;
}
end_execution_reason_t parse_execution_context_t::run_job_list(
const ast::andor_job_list_t &job_list_node, const block_t *associated_block) {
auto result = end_execution_reason_t::ok;
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 14:42:59 +00:00
for (size_t i = 0; i < job_list_node.count(); i++) {
const ast::andor_job_t *aoj = job_list_node.at(i);
result = test_and_run_1_job_conjunction(aoj->job(), associated_block);
}
// Returns the result of the last job executed or skipped.
return result;
}
end_execution_reason_t parse_execution_context_t::eval_node(const ast::statement_t &statement,
const block_t *associated_block) {
// Note we only expect block-style statements here. No not statements.
enum end_execution_reason_t status = end_execution_reason_t::ok;
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 14:42:59 +00:00
const auto contents = statement.contents().ptr();
if (const auto *block = contents->try_as_block_statement()) {
status = this->run_block_statement(*block, associated_block);
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 14:42:59 +00:00
} else if (const auto *ifstat = contents->try_as_if_statement()) {
status = this->run_if_statement(*ifstat, associated_block);
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 14:42:59 +00:00
} else if (const auto *switchstat = contents->try_as_switch_statement()) {
status = this->run_switch_statement(*switchstat);
} else {
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 14:42:59 +00:00
FLOGF(error, L"Unexpected node %ls found in %s", statement.describe()->c_str(),
2019-06-04 03:30:48 +00:00
__FUNCTION__);
abort();
}
return status;
}
end_execution_reason_t parse_execution_context_t::eval_node(const ast::job_list_t &job_list,
const block_t *associated_block) {
assert(associated_block && "Null block");
// Check for infinite recursion: a function which immediately calls itself..
wcstring func_name;
if (const auto *infinite_recursive_node =
this->infinite_recursive_statement_in_job_list(job_list, &func_name)) {
// We have an infinite recursion.
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 14:42:59 +00:00
return this->report_error(STATUS_CMD_ERROR, *infinite_recursive_node->ptr(),
INFINITE_FUNC_RECURSION_ERR_MSG, func_name.c_str());
}
2022-10-29 03:07:36 +00:00
// Check for stack overflow in case of function calls (regular stack overflow) or string
// substitution blocks, which can be recursively called with eval (issue #9302).
if ((associated_block->type() == block_type_t::top &&
parser->function_stack_is_overflowing()) ||
(associated_block->type() == block_type_t::subst && parser->is_eval_depth_exceeded())) {
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 14:42:59 +00:00
return this->report_error(STATUS_CMD_ERROR, *job_list.ptr(),
CALL_STACK_LIMIT_EXCEEDED_ERR_MSG);
}
return this->run_job_list(job_list, associated_block);
}
void parse_execution_context_t::setup_group(job_t *j) {
// We can use the parent group if it's compatible and we're not backgrounded.
if (ctx.job_group && (ctx.job_group->has_job_id() || !j->wants_job_id()) &&
!j->is_initially_background()) {
j->group = ctx.job_group;
return;
}
if (j->processes.front()->is_internal() || !this->use_job_control()) {
// This job either doesn't have a pgroup (e.g. a simple block), or lives in fish's pgroup.
rust::Box<job_group_t> group = create_job_group_ffi(j->command(), j->wants_job_id());
j->group = box_to_shared_ptr(std::move(group));
} else {
// This is a "real job" that gets its own pgroup.
j->processes.front()->leads_pgrp = true;
bool wants_terminal = !parser->libdata().is_event;
auto group = create_job_group_with_job_control_ffi(j->command(), wants_terminal);
j->group = box_to_shared_ptr(std::move(group));
}
j->group->set_is_foreground(!j->is_initially_background());
j->mut_flags().is_group_root = true;
}
bool parse_execution_context_t::use_job_control() const {
if (parser->is_command_substitution()) {
return false;
}
switch (get_job_control_mode()) {
case job_control_t::all:
return true;
case job_control_t::interactive:
return parser->is_interactive();
case job_control_t::none:
return false;
}
DIE("Unreachable");
}
int parse_execution_context_t::line_offset_of_node(const ast::job_pipeline_t *node) {
// If we're not executing anything, return -1.
if (!node) {
return -1;
}
// If for some reason we're executing a node without source, return -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 14:42:59 +00:00
if (!node->try_source_range()) {
return -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 14:42:59 +00:00
return this->line_offset_of_character_at_offset(node->source_range().start);
}
2014-03-31 17:01:39 +00:00
int parse_execution_context_t::line_offset_of_character_at_offset(size_t offset) {
// Count the number of newlines, leveraging our cache.
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 14:42:59 +00:00
assert(offset <= pstree->src().size());
// Easy hack to handle 0.
if (offset == 0) {
return 0;
}
2014-03-31 17:01:39 +00:00
// We want to return (one plus) the number of newlines at offsets less than the given offset.
// cached_lineno_count is the number of newlines at indexes less than cached_lineno_offset.
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 14:42:59 +00:00
const wcstring &str = pstree->src();
if (offset > cached_lineno_offset) {
size_t i;
for (i = cached_lineno_offset; i < offset && str[i] != L'\0'; i++) {
// Add one for every newline we find in the range [cached_lineno_offset, offset).
if (str[i] == L'\n') {
cached_lineno_count++;
}
}
cached_lineno_offset =
i; // note: i, not offset, in case offset is beyond the length of the string
} else if (offset < cached_lineno_offset) {
// Subtract one for every newline we find in the range [offset, cached_lineno_offset).
for (size_t i = offset; i < cached_lineno_offset; i++) {
if (str[i] == L'\n') {
cached_lineno_count--;
}
}
cached_lineno_offset = offset;
}
return cached_lineno_count;
}
int parse_execution_context_t::get_current_line_number() {
int line_number = -1;
int line_offset = this->line_offset_of_node(this->executing_job_node);
if (line_offset >= 0) {
// The offset is 0 based; the number is 1 based.
line_number = line_offset + 1;
}
return line_number;
}
int parse_execution_context_t::get_current_source_offset() const {
int result = -1;
if (executing_job_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 14:42:59 +00:00
if (executing_job_node->try_source_range()) {
result = static_cast<int>(executing_job_node->source_range().start);
}
}
return result;
}