9f09930834
# Description Dividing two ints can currently return either an int or a float. Not having a single return type for an operation between two types seems problematic. Additionally, the type signature for division says that dividing two ints returns only an int which does not match the current implementation (it can also return a float). This PR changes division between almost all types to return a float (except for `filesize / number` or `duration / number`, since there are no float representations for these types). Currently, floor division between certain types is not implemented even though the type signature allows it. Also, the current implementation of floor division uses a combination of clamping and flooring rather than simply performing floor division which this PR fixes. Additionally, the signature was changed so that `int // float`, `float // int`, and `float // float` now return float instead of int. This matches the automatic float promotion in the rest of the operators (as well as how Python does floor division which I think is the original inspiration). Since regular division has always returned fractional values (and now returns a float to reflect that), `mod` is now defined in terms of floor division. That is, `D // d = q`, `D mod d = r`, and `D = d * q + r `. This is just like the `%` operator in Python, which is also based off floor division (at least for ints and floats). Additionally, implementations missing from `mod`'s current type signature have been added (`duration mod int` and `duration mod float`). This PR also overhauls the overflow checking and errors for div, mod, and floor div. If an operation overflows, it will now cause an error. # User-Facing Changes - Div now returns a float in most cases. - Floor division now actually does floor division. - Floor division now does automatic float promotion, returning a float in more instances. - Floor division now actually allows division with filesize and durations as its type signature claimed. - Mod is now defined and implemented in terms of floor division rather than truncating division. - Mod now actually allows filesize and durations as its type signature claimed. - Div, mod, and floor div now all have proper overflow checks. ## Examples When the divisor and the dividend have the same sign, the quotient and remainder will be the same as before. (Except that this PR will give more accurate results, since it does not do an intermediate float conversion). If the signs of the divisor and dividend are different, then the results will be different, or rather actually correct. Before: ```nu let q = 8 // -3 # -3 let r = 8 mod -3 # 2 8 == $q * -3 + $r # false ``` After: ```nu let q = 8 // -3 # -3 let r = 8 mod -3 # -1 8 == $q * -3 + $r # true ``` Before: ```nu let q = -8 // 3 # -3 let r = -8 mod 3 # -2 -8 == $q * 3 + $r # false ``` After: ```nu let q = -8 // 3 # -3 let r = -8 mod 3 # 1 -8 == $q * 3 + $r # true ``` # Tests + Formatting Added a few tests. # After Submitting Probably update the docs. |
||
---|---|---|
.. | ||
fuzz | ||
src | ||
tests | ||
Cargo.toml | ||
LICENSE | ||
README.md |
nu-parser, the Nushell parser
Nushell's parser is a type-directed parser, meaning that the parser will use type information available during parse time to configure the parser. This allows it to handle a broader range of techniques to handle the arguments of a command.
Nushell's base language is whitespace-separated tokens with the command (Nushell's term for a function) name in the head position:
head1 arg1 arg2 | head2
Lexing
The first job of the parser is to a lexical analysis to find where the tokens start and end in the input. This turns the above into:
<item: "head1">, <item: "arg1">, <item: "arg2">, <pipe>, <item: "head2">
At this point, the parser has little to no understanding of the shape of the command or how to parse its arguments.
Lite parsing
As Nushell is a language of pipelines, pipes form a key role in both separating commands from each other as well as denoting the flow of information between commands. The lite parse phase, as the name suggests, helps to group the lexed tokens into units.
The above tokens are converted the following during the lite parse phase:
Pipeline:
Command #1:
<item: "head1">, <item: "arg1">, <item: "arg2">
Command #2:
<item: "head2">
Parsing
The real magic begins to happen when the parse moves on to the parsing stage. At this point, it traverses the lite parse tree and for each command makes a decision:
- If the command looks like an internal/external command literal: e.g.
foo
or/usr/bin/ls
, it parses it as an internal or external command - Otherwise, it parses the command as part of a mathematical expression
Types/shapes
Each command has a shape assigned to each of the arguments it reads in. These shapes help define how the parser will handle the parse.
For example, if the command is written as:
where $x > 10
When the parsing happens, the parser will look up the where
command and find its Signature. The Signature states what flags are allowed and what positional arguments are allowed (both required and optional). Each argument comes with a Shape that defines how to parse values to get that position.
In the above example, if the Signature of where
said that it took three String values, the result would be:
CallInfo:
Name: `where`
Args:
Expression($x), a String
Expression(>), a String
Expression(10), a String
Or, the Signature could state that it takes in three positional arguments: a Variable, an Operator, and a Number, which would give:
CallInfo:
Name: `where`
Args:
Expression($x), a Variable
Expression(>), an Operator
Expression(10), a Number
Note that in this case, each would be checked at compile time to confirm that the expression has the shape requested. For example, "foo"
would fail to parse as a Number.
Finally, some Shapes can consume more than one token. In the above, if the where
command stated it took in a single required argument, and that the Shape of this argument was a MathExpression, then the parser would treat the remaining tokens as part of the math expression.
CallInfo:
Name: `where`
Args:
MathExpression:
Op: >
LHS: Expression($x)
RHS: Expression(10)
When the command runs, it will now be able to evaluate the whole math expression as a single step rather than doing any additional parsing to understand the relationship between the parameters.
Making space
As some Shapes can consume multiple tokens, it's important that the parser allow for multiple Shapes to coexist as peacefully as possible.
The simplest way it does this is to ensure there is at least one token for each required parameter. If the Signature of the command says that it takes a MathExpression and a Number as two required arguments, then the parser will stop the math parser one token short. This allows the second Shape to consume the final token.
Another way that the parser makes space is to look for Keyword shapes in the Signature. A Keyword is a word that's special to this command. For example in the if
command, else
is a keyword. When it is found in the arguments, the parser will use it as a signpost for where to make space for each Shape. The tokens leading up to the else
will then feed into the parts of the Signature before the else
, and the tokens following are consumed by the else
and the Shapes that follow.