DSLX is a domain specific, dataflow-oriented functional language used to build hardware that can also run effectively as host software. Within the XLS project, DSLX is also referred to as "the DSL". The DSL targets the XLS compiler (via conversion to XLS IR) to enable flows for FPGAs and ASICs.
DSLX mimics Rust, while being an immutable expression-based dataflow DSL with hardware-oriented features; e.g. arbitrary bitwidths, entirely fixed size objects, fully analyzeable call graph, etc. To avoid arbitrary new syntax/semantics choices, the DSL mimics Rust where it is reasonably possible; for example, integer conversions all follow the same semantics as Rust.
Note: There are some unnecessary differences today from Rust syntax due to early experimentation, but they are quickly being removed to converge on Rust syntax.
Note that other frontends to XLS core functionality will become available in the future; e.g. xlscc, for users familiar with the C++-and-pragma style of HLS computation. XLS team develops the DSL as part of the XLS project because we believe it can offer significant advantages over the C++-with-pragmas approach.
Dataflow DSLs are a good fit for describing hardware, compared to languages whose design assumes von Neumann style computation (global mutable state, sequential mutation by a sequential thread of control). Using a Domain Specific Language (DSL) provides a more hardware-oriented representation of a given computation that matches XLS compiler (IR) constructs closely. The DSL also allows an exploration of HLS without being encumbered by C++ language or compiler limitations such as non-portable pragmas, magic macros, or semantically important syntactic conventions. The language is still experimental and likely to change, but it is already useful for experimentation and exploration.
This document provides a reference for DSLX, mostly by example. Before perusing it in detail, we recommend you first read the DSLX tutorials to understand the broad strokes of the language.
In this document we use the function to compute a CRC32 checksum to describe
language features. The full code is in
examples/dslx_intro/crc32_one_byte.x.
Just as in languages like Rust/C++, comments start with // and last through
the end of the line.
All identifiers, eg., for function names, parameters, and values, follow the typical naming rules of other languages. The identifiers can start with a character or an underscore, and can then contain more characters, underscores, or numbers. Valid examples are:
a // valid
CamelCase // valid
like_under_scores // valid
__also_ok // valid
_Ok123_321 // valid
_ // valid
2ab // not valid
&ade // not valid
However, we suggest the following DSLX style rules, which mirror the Rust naming conventions.
-
Functions are
written_like_this -
User-defined data types are
NamesLikeThis -
Constant bindings are
NAMES_LIKE_THIS -
_is the "black hole" identifier -- a name that you can bind to but should never read from, akin to Rust's wildcard pattern match or Python's "unused identifier" convention. It should never be referred to in an expression except as a "sink".
NOTE Since mutable locals are not supported, there is also
support for "tick identifiers",
where a ' character may appear anywhere after the first character of an
identifier to indicate "prime"; e.g. let state' = update(state);. By
convention ticks usually come at the end of an identifier. Since this is not
part of Rust's syntax, it is considered experimental at this time.
If you bind a name and do not use it, a warning will be flagged, and warnings are errors by default; e.g. this will flag an unused warning:
#[test]
fn my_test() {
let x = u32:42; // Not used!
}
For cases where it's more readable to keep a name, even though it's unused, you can prefix the name with a leading underscore, like so:
#[test]
fn my_test() {
let (thing_one, _thing_two) = open_crate();
assert_eq(thing_one, u32:1);
}
Note that _thing_two is unused, but a warning is not flagged because we
indicated via a leading underscore that it was ok for the variable to go unused,
because we felt it enhanced readability.
Function definitions begin with the keyword fn, followed by the function name,
a parameter list to the function in parenthesis, followed by an -> and the
return type of the function. After this, curly braces denote the begin and end
of the function body.
The list of parameters can be empty.
A single input file can contain many functions.
Simple examples:
fn ret3() -> u32 {
u32:3 // This function always returns 3.
}
fn add1(x: u32) -> u32 {
x + u32:1 // Returns x + 1, but you knew that!
}
Functions return the result of their last computed expression as their return value. There are no explicit return statements. By implication, functions return exactly one expression; they can't return multiple expressions (but this may change in the future as we migrate towards some Rust semantics).
Tuples should be returned if a function needs to return multiple values.
Parameters are written as pairs name followed by a colon : followed by the
type of that parameter. Each parameter needs to declare its own type.
Examples:
// a simple parameter x of type u32
x: u32
// t is a tuple with 2 elements.
// the 1st element is of type u32
// the 2nd element is a tuple with 3 elements
// the 1st element is of type u8
// the 2nd element is another tuple with 1 element of type u16
// the 3rd element is of type u8
t: (u32, (u8, (u16,), u8))
DSLX functions can be parameterized in terms of the types of its arguments and in terms of types derived from other parametric values. For instance:
fn double(n: u32) -> u32 {
n * u32:2
}
fn self_append<A: u32, B: u32 = {double(A)}>(x: bits[A]) -> bits[B] {
x++x
}
fn main() -> bits[10] {
self_append(u5:1)
}
In self_append(bits[5]:1), we see that A = 5 based off of formal argument
instantiation. Using that value, we can evaluate B = double(A=5). This derived
expression is analogous to C++'s constexpr – a simple expression that can be
evaluated at that point in compilation. Note that the expression must be wrapped
in {} curly braces.
See advanced understanding for more information on parametricity.
In some cases, parametric values cannot be inferred from function arguments,
such as in the
explicit_parametric_simple.x
test:
fn add_one<E:u32, F:u32, G:u32 = E+F>(lhs: bits[E]) -> bits[G] { ... }
For this call to instantiable, both E and F must be specified. Since F
can't be inferred from an argument, we must rely on explicit parametrics:
add_one<u32:1, {u32:2 + u32:3}>(u1:1);
This invocation will bind 1 to E, 5 to F, and 6 to G. Note the curly
braces around the expression-defined parametric: simple literals and constant
references do not need braces (but they can have them), but any other
expression requires them.
Without curly braces, explicit parametric expressions could be ambiguous; consider the following, slightly changed from the previous example:
add_one<u32:1, u32:2>(u32:3)>(u1:1);
Is the statement above computing add_one<1, (2 > 3)>(1), or is it computing
(add_one<1, 2>(3)) > 1)? Without additional (and subtle and perhaps surprising)
contextual precedence rules, this would be ambiguous and could lead to a parse
error or, even worse, unexpected behavior.
Fortunately, we can look to Rust for inspiration. Rust's const generics RPF
introduced the { } syntax for disambiguating just this case in generic
specifications. With this, any expressions present in a parametric specification
must be contained within curly braces, as in the original example.
At present, if the braces are omitted, some unpredictable error will occur. Work to improve this is tracked in XLS GitHub issue #321.
Function calls are expressions and look and feel just like one would expect from other languages. For example:
fn callee(x: bits[32], y: bits[32]) -> bits[32] {
x + y
}
fn caller() -> u32 {
callee(u32:2, u32:3)
}
If more than one value should be returned by a function, a tuple type should be returned.
The DSL has several built-in functions and standard library modules. For details on the available functions to invoke, see DSLX Built-In Functions and Standard Library.
The most fundamental type in DSLX is a variable length bit type denoted as
bits[n], where n is a constant. For example:
bits[1] // a single bit
uN[1] // explicitly noting single bit is unsigned
u1 // convenient shorthand for bits[1]
bits[8] // an 8-bit datatype, yes, a byte
u8 // convenient shorthand for bits[8]
bits[32] // a 32-bit datatype
u32 // convenient shorthand for bits[32]
bits[256] // a 256-bit datatype
bits[0] // possible, but, don't do that
DSLX introduces aliases for commonly used types, such as u8 for an 8-wide bit
type, or u32 for a 32-bit wide bit type. These are defined up to u64.
All u*, uN[*], and bits[*] types are interpreted as unsigned integers.
Signed integers are specified via s* and sN[*]. Similarly to unsigned
numbers, the s* shorthands are defined up to s64. For example:
sN[1]
s1
sN[64]
s64
sN[256]
sN[0]
s0
Signed numbers differ in their behavior from unsigned numbers primarily via operations like comparisons, (variable width) multiplications, and divisions.
Bit types have helpful type-level attributes that provide limit values, similar
to std::numeric_limits in C++. For example:
u3::MAX // u3:0b111 the "fill with ones" value
s3::MAX // s3:0b011 the "maximum signed" value
u3::ZERO // u3:0b000 the "fill with zeros" value
s3::ZERO // s3:0b000 the "fill with zeros" value
Characters are a special case of bits types: they are implicitly-type as u8. Characters can be used just as traditional bits:
fn add_to_null(input: u8) -> u8 {
let null:u8 = '\0';
input + null
}
#[test]
fn test_main() {
assert_eq('a', add_to_null('a'))
}
DSLX character constants support the full Rust set of escape sequences with the exception of unicode.
DSLX supports enumerations as a way of defining a group of related, scoped, named constants that do not pollute the module namespace. For example:
enum Opcode : u3 {
FIRE_THE_MISSILES = 0,
BE_TIRED = 1,
TAKE_A_NAP = 2,
}
fn get_my_favorite_opcode() -> Opcode {
Opcode::FIRE_THE_MISSILES
}
Note the use of the double-colon to reference the enum value. This code
specifies that the enum behaves like a u3: its storage and extension (via
casting) behavior are defined to be those of a u3. Attempts to define an enum
value outside of the representable u3 range will produce a compile time error.
enum Opcode : u3 {
FOO = 8 // Causes compile time error!
}
Enums can be compared for equality/inequality, but they do not permit arithmetic operations, they must be cast to numerical types in order to perform arithmetic:
enum Opcode: u3 {
NOP = 0,
ADD = 1,
SUB = 2,
MUL = 3,
}
fn same_opcode(x: Opcode, y: Opcode) -> bool {
x == y // ok
}
fn next_in_sequence(x: Opcode, y: Opcode) -> bool {
// x+1 == y // does not work, arithmetic!
x as u3 + u3:1 == (y as u3) // ok, casted first
}
As mentioned above, casting of enum-values works with the same casting/extension rules that apply to the underlying enum type definition. For example, this cast will sign extend because the source type for the enum is signed. (See numerical conversions for the full description of extension/truncation behavior.)
enum MySignedEnum : s3 {
LOW = -1,
ZERO = 0,
HIGH = 1,
}
fn extend_to_32b(x: MySignedEnum) -> u32 {
x as u32 // Sign-extends because the source type is signed.
}
#[test]
fn test_extend_to_32b() {
assert_eq(extend_to_32b(MySignedEnum::LOW), u32:0xffffffff)
}
Casting to an enum is also permitted. However, in most cases errors from invalid casting can only be found at runtime, e.g., in the DSL interpreter or flagging a fatal error from hardware. Because of that, it is recommended to avoid such casts as much as possible.
A tuple is a fixed-size ordered set, containing elements of heterogeneous types. Tuples elements can be any type, e.g. bits, arrays, structs, tuples. Tuples may be empty (an empty tuple is also known as the unit type), or contain one or more types.
Examples of tuple values:
// The unit type, carries no information.
let unit = ();
// A tuple containing two bits-typed elements.
let pair = (u3:0b100, u4:0b1101);
Example of a tuple type:
// The type of a tuple with 2 elements.
// the 1st element is of type u32
// the 2nd element is a tuple with 3 elements
// the 1st element is of type u8
// the 2nd element is another tuple with 1 element of type u16
// the 3rd element is of type u8
type MyTuple = (u32, (u8, (u16,), u8));
To access individual tuple elements use simple indices, starting at 0. For example, to access the second element of a tuple (index 1):
#[test]
fn test_tuple_access() {
let t = (u32:2, u8:3);
assert_eq(u8:3, t.1)
}
Such indices can only be numeric literals; parametric symbols are not allowed.
Tuples can be "destructured", similarly to how pattern matching works in match
expressions, which provides a convenient syntax to name elements of a tuple for
subsequent use. See a and b in the following:
#[test]
fn test_tuple_destructure() {
let t = (u32:2, u8:3);
let (a, b) = t;
assert_eq(u32:2, a);
assert_eq(u8:3, b)
}
Just as values can be discarded in a let by using the "black hole identifier"
_, don't-care values can also be discarded when destructuring a tuple:
#[test]
fn test_black_hole() {
let t = (u32:2, u8:3, true);
let (_, _, v) = t;
assert_eq(v, true)
}
Structures are similar to tuples, but provide two additional capabilities: we name the slots (i.e. struct fields have names while tuple elements only have positions), and we introduce a new type.
The following syntax is used to define a struct:
struct Point {
x: u32,
y: u32
}
Once a struct is defined it can be constructed by naming the fields in any order:
struct Point {
x: u32,
y: u32,
}
#[test]
fn test_struct_equality() {
let p0 = Point { x: u32:42, y: u32:64 };
let p1 = Point { y: u32:64, x: u32:42 };
assert_eq(p0, p1)
}
There is a simple syntax when creating a struct whose field names match the names of in-scope values:
struct Point { x: u32, y: u32, }
#[test]
fn test_struct_equality() {
let x = u32:42;
let y = u32:64;
let p0 = Point { x, y };
let p1 = Point { y, x };
assert_eq(p0, p1)
}
Struct fields can also be accessed with "dot" syntax:
struct Point {
x: u32,
y: u32,
}
fn f(p: Point) -> u32 {
p.x + p.y
}
fn main() -> u32 {
f(Point { x: u32:42, y: u32:64 })
}
#[test]
fn test_main() {
assert_eq(u32:106, main())
}
Note that structs cannot be mutated "in place", the user must construct new values by extracting the fields of the original struct mixed together with new field values, as in the following:
struct Point3 {
x: u32,
y: u32,
z: u32,
}
fn update_y(p: Point3, new_y: u32) -> Point3 {
Point3 { x: p.x, y: new_y, z: p.z }
}
fn main() -> Point3 {
let p = Point3 { x: u32:42, y: u32:64, z: u32:256 };
update_y(p, u32:128)
}
#[test]
fn test_main() {
let want = Point3 { x: u32:42, y: u32:128, z: u32:256 };
assert_eq(want, main())
}
The DSL has syntax for conveniently producing a new value with a subset of fields updated to reduce verbosity. The "struct update" syntax is:
struct Point3 {
x: u32,
y: u32,
z: u32,
}
fn update_y(p: Point3) -> Point3 {
Point3 { y: u32:42, ..p }
}
fn update_x_and_y(p: Point3) -> Point3 {
Point3 { x: u32:42, y: u32:42, ..p }
}
DSLX also supports parametric structs. For more information on how type-parametricity works, see the parametric functions section.
fn double(n: u32) -> u32 { n * u32:2 }
struct Point<N: u32, M: u32 = {double(N)}> {
x: bits[N],
y: bits[M],
}
fn make_point<A: u32, B: u32>(x: bits[A], y: bits[B]) -> Point<A, B> {
Point{ x, y }
}
#[test]
fn test_struct_construction() {
let p = make_point(u16:42, u32:42);
assert_eq(u16:42, p.x)
}
As mentioned above, a struct definition introduces a new type. Structs are nominally typed, as opposed to structurally typed (note that tuples are structurally typed). This means that structs with different names have different types, regardless of whether those structs have the same structure (i.e. even when all the fields of two structures are identical, those structures are a different type when they have a different name).
struct Point {
x: u32,
y: u32,
}
struct Coordinate {
x: u32,
y: u32,
}
fn f(p: Point) -> u32 {
p.x + p.y
}
#[test]
fn test_ok() {
assert_eq(f(Point { x: u32:42, y: u32:64 }), u32:106)
}
#[test]
fn test_type_checker_error() {
assert_eq(f(Coordinate { x: u32:42, y: u32:64 }), u32:106)
}
Arrays can be constructed via bracket notation. All values that make up the
array must have the same type. Arrays can be indexed with indexing notation
(a[i]) to retrieve a single element.
fn main(a: u32[2], i: u1) -> u32 {
a[i]
}
#[test]
fn test_main() {
let x = u32:42;
let y = u32:64;
// Make an array with "bracket notation".
let my_array: u32[2] = [x, y];
assert_eq(main(my_array, u1:0), x);
assert_eq(main(my_array, u1:1), y);
}
Because arrays with repeated trailing elements are common, the DSL supports
ellipsis (...) at the end of an array to fill the remainder of the array with
the last noted element. Because the compiler must know how many elements to
fill, in order to use the ellipsis the type must be annotated explicitly as
shown.
fn make_array(x: u32) -> u32[3] {
u32[3]:[u32:42, x, ...]
}
#[test]
fn test_make_array() {
assert_eq(u32[3]:[u32:42, u32:42, u32:42], make_array(u32:42));
assert_eq(u32[3]:[u32:42, u32:64, u32:64], make_array(u32:64));
}
Note google/xls#917: arrays with length zero will typecheck, but fail to work in most circumstances. Eventually, XLS should support them but they can't be used currently.
TODO(meheff): Explain arrays and the intricacies of our bits type interpretation and how it affects arrays of bits etc.
Character strings are a special case of array types, being implicitly-sized arrays of u8 elements. String constants can be used just as traditional arrays:
fn add_one<N: u32>(input: u8[N]) -> u8[N] {
for (i, result) : (u32, u8[N]) in u32:0..N {
update(result, i, result[i] + u8:1)
}(input)
}
#[test]
fn test_main() {
assert_eq("bcdef", add_one("abcde"))
}
DSLX string constants support the
full Rust set of escape sequences -
note that unicode escapes get encoded to their UTF-8 byte sequence. In other
words, the sequence \u{10CB2F} will result in an array with hexadecimal values
F4 8C AC AF.
Moreover, string can be composed of characters.
fn string_composed_characters() -> u8[10] {
['X', 'L', 'S', ' ', 'r', 'o', 'c', 'k', 's', '!']
}
#[test]
fn test_main() {
assert_eq("XLS rocks!", string_composed_characters())
}
DLSX supports the definition of type aliases.
Type aliases can be used to provide a more human-readable name for an existing type. The new name is on the left, the existing name on the right:
type Weight = u6;
We can create an alias for an imported type:
// Note: this imports an external file in the codebase.
import xls.dslx.tests.mod_imported
type MyEnum = mod_imported::MyEnum;
fn main(x: u8) -> MyEnum {
x as MyEnum
}
#[test]
fn test_main() {
assert_eq(main(u8:42), MyEnum::FOO);
assert_eq(main(u8:64), MyEnum::BAR);
}
Type aliases can also provide a descriptive name for a tuple type (which is otherwise anonymous). For example, to define a tuple type that represents a float number with a sign bit, an 8-bit mantissa, and a 23-bit mantissa, one would write:
type F32 = (u1, u8, u23);
After this definition, the F32 may be used as a type annotation
interchangeably with (u1, u8, u23).
Note, however, that structs are generally preferred for this purpose, as they are more readable and users do not need to rely on tuple elements having a stable order in the future (i.e., they are resilient to refactoring).
Bit types can be cast from one bit-width to another with the as keyword. Types
can be widened (increasing bit-width), narrowed (decreasing bit-width) and/or
changed between signed and unsigned. Some examples are found below. See
Numerical Conversions for a description of the
semantics.
#[test]
fn test_narrow_cast() {
let twelve = u4:0b1100;
assert_eq(twelve as u2, u2:0)
}
#[test]
fn test_widen_cast() {
let three = u2:0b11;
assert_eq(three as u4, u4:3)
}
#[test]
fn test_narrow_signed_cast() {
let negative_seven = s4:0b1001;
assert_eq(negative_seven as u2, u2:1)
}
#[test]
fn test_widen_signed_cast() {
let negative_one = s2:0b11;
assert_eq(negative_one as s4, s4:-1)
}
#[test]
fn test_widen_to_unsigned() {
let negative_one = s2:0b11;
assert_eq(negative_one as u3, u3:0b111)
}
#[test]
fn test_widen_to_signed() {
let three = u2:0b11;
assert_eq(three as u3, u3:0b011)
}
DSLX performs type checking and produces an error if types in an expression don't match up.
let expressions also perform type inference, which is quite convenient. For
example, instead of writing:
let ch: u32 = (e & f) ^ ((!e) & g);
let (h, g, f): (u32, u32, u32) = (g, f, e);
one can write the following, as long as the types can be properly inferred:
let ch = (e & f) ^ ((!e) & g);
let (h, g, f) = (g, f, e);
Note that type annotations can still be added and be used for program understanding, as they they will be checked by DSLX.
All expressions in the language's expression grammar have a deductive type inference rule. The types must be known for inputs to an operator/function1 and every expression has a way to determine its type from its operand expressions.
DSLX uses deductive type inference to check the types present in the program. Deductive type inference is a set of (typically straight-forward) deduction rules: Hindley-Milner style deductive type inference determines the result type of a function with a rule that only observes the input types to that function. (Note that operators like '+' are just slightly special functions in that they have pre-defined special-syntax-rule names.)
In DSLX code, the "environment" where names are bound (sometimes also referred
to as a symbol table) is called the
Bindings -- it
maps identifiers to the AST node that defines the name ({string: AstNode}),
which can be combined with a mapping from AST node to its deduced type
({AstNode: ConcreteType}) to resolve the type of an identifier in the program.
Let is one of the key nodes that populates these Bindings, but anything that
creates a bound name does as well (e.g. parameters, for loop induction
variables, etc.).
For example, consider the binary (meaning takes two operands) / infix (meaning it syntactically is placed in the center of its operands) '+' operator. The simple deductive type inference rule for '+' is:
(T, T) -> T
Meaning that the left hand side operand to the '+' operator is of some type (call it T), the right hand side operand to the '+' operator must be of that same type, T, and the result of that operator is then (deduced) to be of the same type as its operands, T.
Let's instantiate this rule in a function:
fn add_wrapper(x: bits[2], y: bits[2]) -> bits[2] {
x + y
}
This function wraps the '+' operator. It presents two arguments to the '+'
operator and then checks that the annotated return type on add_wrapper matches
the deduced type for the body of that function; that is, we ask the following
question of the '+' operator (since the type of the operands must be known at
the point the add is performed):
(bits[2], bits[2]) -> ?
To resolve the '?' the following procedure is being used:
- Pattern match the rule given above
(T, T) -> Tto determine the type T: the left hand side operand isbits[2], called T. - Check that the right hand side is also that same T, which it is: another
bits[2]. - Deduce that the result type is that same type T:
bits[2]. - That becomes the return type of the body of the function. Check that it is the same type as the annotated return type for the function, and it is!
The function is annotated to return bits[2], and the deduced type of the body
is also bits[2]. Qed.
A type error would occur in the following:
fn add_wrapper(x: bits[2], y: bits[3]) -> bits[2] {
x + y
}
Applying the type deduction rule for '+' finds an inconsistency. The left hand
side operand has type bits[2], called T, but the right hand side is bits[3],
which is not the same as T. Because the deductive type inference rule does not
say what to do when the operand types are different, it results in a type error
which is flagged at this point in the program.
In the DSL, let is an expression. It may not seem obvious at a glance, but it
is! As a primer see the type inference background
and how names are resolved in an environment.
"let" expressions are of the (Rust-inspired) form:
let $name: $annotated_type = $expr; $subexpr
$name gets "bound" to a value of type $annotated_type. The let typecheck
rule must both check that $expr is of type $annotated_type, as well as
determine the type of $subexpr, which is the type of the overall "let
expression".
In this example, the result of the let expression is the return value --
$subexpr (x+x) can use the $name (x) which was "bound":
fn main(y: u32) -> u64 {
let x: u64 = y as u64;
x+x
}
If we invoke main(u32:2) we will the evaluate let expression -- it creates a
binding of x to the value u64:2, and then evaluates the expression x+x in
that environment, so the result of the let expression's $subexpr is u64:4.
DSLX modules can import other modules via the import keyword. Circular imports
are not permitted (the dependencies among DSLX modules must form a DAG, as in
languages like Go).
The import statement takes the following form (note the lack of semicolon):
import path.to.my.imported_module
With that statement, the module will be accessible as (the trailing identifier
after the last dot) imported_module; e.g. the program can refer to
imported_module::IMPORTED_MODULE_PUBLIC_CONSTANT.
NOTE Imports are relative to the Bazel "depot root" -- for external use of the
tools a DSLX_PATH will be exposed, akin to a PYTHONPATH, for users to
indicate paths where were should attempt module discovery.
NOTE Importing does not introduce any names into the current file other than
the one referred to by the import statement. That is, if imported_module had a
constant defined in it FOO, this is referred to via imported_module::FOO,
FOO does not "magically" get put in the current scope. This is analogous to
how wildcard imports are discouraged in other languages (e.g. from import * in
Python) on account of leading to "namespace pollution" and needing to specify
what happens when names conflict.
If you want to change the name of the imported module (for reference inside of
the importing file) you can use the as keyword:
import path.to.my.imported_module as im
Just using the above construct,
imported_module::IMPORTED_MODULE_PUBLIC_CONSTANT is not valid, only
im::IMPORTED_MODULE_PUBLIC_CONSTANT. However, both statements can be used on
different lines:
import path.to.my.imported_module
import path.to.my.imported_module as im
In this case, either im::IMPORTED_MODULE_PUBLIC_CONSTANT or
imported_module::IMPORTED_MODULE_PUBLIC_CONSTANT can be used to refer to the
same thing.
Here is an example using the same function via two different aliases for the same module:
// Note: this imports an external file in the codebase under two different
// names.
import xls.dslx.tests.mod_imported
import xls.dslx.tests.mod_imported as mi
fn main(x: u3) -> u1 {
mod_imported::my_lsb(x) || mi::my_lsb(x)
}
#[test]
fn test_main() {
assert_eq(u1:0b1, main(u3:0b001))
}
Module members are private by default and not accessible from any importing
module. To make a member public/visible to importing modules, the pub keyword
must be added as a prefix; e.g.
const FOO = u32:42; // Not accessible to importing modules.
pub const BAR = u32:64; // Accessible to importing modules.
This applies to other things defined at module scope as well: functions, enums, type aliases, etc.
import xls.dslx.tests.mod_imported
import xls.dslx.tests.mod_imported as mi
fn main(x: u3) -> u1 {
mod_imported::my_lsb(x) || mi::my_lsb(x)
}
#[test]
fn test_main() {
assert_eq(u1:0b1, main(u3:0b001))
}
The const keyword is used to define module-level constant values. Named
constants should be usable anywhere a literal value can be used:
const FOO = u8:42;
fn match_const(x: u8) -> u8 {
match x {
FOO => u8:0,
_ => u8:42,
}
}
#[test]
fn test_match_const_not_binding() {
assert_eq(u8:42, match_const(u8:0));
assert_eq(u8:42, match_const(u8:1));
assert_eq(u8:0, match_const(u8:42));
}
fn h(t: (u8, (u16, u32))) -> u32 {
match t {
(FOO, (x, y)) => (x as u32) + y,
(_, (y, u32:42)) => y as u32,
_ => u32:7,
}
}
#[test]
fn test_match_nested() {
assert_eq(u32:3, h((u8:42, (u16:1, u32:2))));
assert_eq(u32:1, h((u8:0, (u16:1, u32:42))));
assert_eq(u32:7, h((u8:0, (u16:1, u32:0))));
}
DSLX supports construction of literals using the syntax Type:Value. For
example u16:1 is a 16-wide bit array with its least significant bit set to
one. Similarly s8:12 is an 8-wide bit array with its least significant four
bits set to 1100.
DSLX supports initializing using binary, hex or decimal syntax. So
#[test]
fn test_literal_initialization() {
assert_eq(u8:12, u8:0b00001100);
assert_eq(u8:12, u8:0x0c);
}
When constructing literals DSLX will trigger an error if the constant will not
fit in a bit array of the annotated sized, so for example trying to construct
the literal u8:256 will trigger an error of the form:
TypeInferenceError: uN[8] Value '256' does not fit in the bitwidth of a uN[8] (8)
But what about s8:128 ? This is a valid literal, even though a signed 8-bit
integer cannot represent it. The following code offers a clue.
#[test]
fn test_signed_literal_initialization() {
assert_eq(s8:128, s8:-128);
assert_eq(s8:128, s8:0b10000000);
}
What is happening here is that, 128 is being used as a bit pattern rather than as the number 128 to initialize the literal. It is only when the bit pattern cannot fit in the width of the literal that an error is triggered.
Note that behaviour is different from Rust, where it will trigger an error, and the fact that DSLX considers this valid may change in the future.
DSLX supports two types of unary expressions with type signature
(xN[N]) -> xN[N]:
- bit-wise not (the
!operator) - negate (the
-operator, computes the two's complement negation)
DSLX supports a familiar set of binary expressions. There are two categories of binary expressions. A category where both operands to the expression must be of the same bit type (i.e., not arrays or tuples), and a category where the operands can be of arbitrary bit types (i.e. shift expressions).
The following expressions have type signature (xN[N], xN[N]) -> xN[N].
- bit-wise or (
|) - bit-wise and (
&) - add (
+) - subtract (
-) - xor (
^) - multiply (
*) - logical or (
||) - logical and (
&&)
Things like
std::smul are
convenient helpers when you are working with mixed widths. Because these
expressions return the same type as the operands, if you want a carry you need
to widen the inputs (e.g.
std::uadd_with_overflow
). The optimizer will narrow the operands and produce efficient hardware,
especially with trivial zero-/sign-extended operands like std::smul and
std::uadd_with_overflow.
Shift expressions include:
- shift-right logical (
>>) - shift-left (
<<)
These are binary operations that don't require the same type on the left and
right hand side. The right hand side must be unsigned, but it does not need to
be the same type or width as the left hand side, i.e. the type signature for
these operations is: (xN[M], uN[N]) -> xN[M]. If the right hand side is a
literal value it does not need to be type annotated. For example:
fn shr_two(x: s32) -> s32 {
x >> 2
}
Note that, as in Rust, the semantics of the shift-right (>>) operation depends
on the signedness of the left hand side. For a signed-type left hand side, the
shift-right (>>) operation performs a shift-right arithmetic and, for a
unsigned-type left hand side, the shift-right (>>) operation performs a
shift-right (logical).
For comparison expressions the types of both operands must match. However these
operations return a result of type bits[1], aka bool.
- equal (
==) - not-equal (
!=) - greater-equal (
>=) - greater (
>) - less-equal (
<=) - less (
<)
Bitwise concatenation is performed with the ++ operator. The value on the left
hand side becomes the most significant bits, the value on the right hand side
becomes the least significant bits. These may be chained together as shown
below:
#[test]
fn test_bits_concat() {
assert_eq(u8:0b11000000, u2:0b11 ++ u6:0b000000);
assert_eq(u8:0b00000111, u2:0b00 ++ u6:0b000111);
assert_eq(u6:0b100111, u1:1 ++ u2:0b00 ++ u3:0b111);
assert_eq(u6:0b001000, u1:0 ++ u2:0b01 ++ u3:0b000);
assert_eq(u32:0xdeadbeef, u16:0xdead ++ u16:0xbeef);
}
Block expressions enable subordinate scopes to be defined, e.g.:
let a = {
let b = u32:1;
b + u32:3
};
The value of a block expression is that of its last contained expression, or (), if a final expression is omitted:
let a = { let b = u32:1; };
In the above case, a is equal to ().
Since DSLX does not currently have the concept of lifetimes, and since names can
be rebound (i.e., there's no concept of mutability, allowing let a = u32:0; let a = u32:1;), blocks are primarily for readability at this time, (side from
their use as the "body" of functions and loops).
Match expressions permit "pattern matching" on data, like a souped-up switch statement. It can both test for values (like a conditional guard) and bind values to identifiers for subsequent use. For example:
fn f(t: (u8, u32)) -> u32 {
match t {
(u8:42, y) => y,
(_, y) => y+u32:77
}
}
If the first member of the tuple is the value is 42, we pass the second tuple
member back as-is from the function. Otherwise, we add 77 to the value and
return that. The _ symbolizes "I don't care about this value".
Just like literal constants, pattern matching can also match via named constants; For example, consider this variation on the above:
const MY_FAVORITE_NUMBER = u8:42;
fn f(t: (u8, u32)) -> u32 {
match t {
(MY_FAVORITE_NUMBER, y) => y,
(_, y) => y+u32:77
}
}
This also works with nested tuples; for example:
const MY_FAVORITE_NUMBER = u8:42;
fn f(t: (u8, (u16, u32))) -> u32 {
match t {
(MY_FAVORITE_NUMBER, (y, z)) => y as u32 + z,
(_, (y, u32:42)) => y as u32,
_ => u32:7
}
}
Here we use a "catch all" wildcard pattern in the last match arm to ensure the match expression always matches the input somehow.
We can also match on ranges of values using the range syntax:
fn f(x: u32) -> u32 {
match x {
u32:1..u32:3 => u32:0,
_ => x
}
}
#[test]
fn test_f() {
assert_eq(f(u32:1), u32:0);
assert_eq(f(u32:2), u32:0);
// Note: the limit of the range syntax is exclusive.
assert_eq(f(u32:3), u32:3);
}
match will flag an error if a syntactically identical pattern is typed
twice; e.g.
const FOO = u32:42;
fn f(x: u32) -> u2 {
match x {
FOO => u2:0,
FOO => u2:1, // Identical pattern!
_ => u2:2,
}
}
Only the first pattern will ever match, so it is fully redundant (and therefore likely a user error they'd like to be informed of). Note that equivalent but not syntactically identical patterns will not be flagged in this way.
const FOO = u32:42;
const BAR = u32:42; // Compares `==` to `FOO`.
fn f(x: u32) -> u2 {
match x {
FOO => u2:0,
BAR => u2:1, // _Equivalent_ pattern, but not syntactically identical.
_ => u2:2,
}
}
let expressions work the same way as let expressions in other functional languages (such as the ML family languages). let expressions provide a nested, lexically-scoped, list of binding definitions. The scope of the binding is the expression on the right hand side of the declaration. For example:
let a: u32 = u32:1 + u32:2;
let b: u32 = a + u32:3;
b
would bind (and return as a value) the value 6 which corresponds to b when
evaluated. In effect there is little difference to other languages like C/C++ or
Python, where the same result would be achieved with code similar to this:
a = 1 + 2
b = a + 3
return bHowever, let expressions are lexically scoped. In above example, the value 3
is bound to a only during the combined let expression sequence. There is no
other type of scoping in DSLX.
DSLX offers an if expression, which is very similar to the Rust if
expression. Blueprint:
if condition { consequent } else { alternate }
This corresponds to the C/C++ ternary ?: operator:
condition ? consequent : alternate
Note: both the if and else are required to be present, as with the ?:
operator, unlike a C++ if statement. This is because it is an expression
that produces a result value, not a statement that causes a mutating effect.
Furthermore, you can have multiple branches via else if:
if condition0 { consequent0 } else if condition1 { consequent1 } else { alternate }
which corresponds to the C/C++:
condition0 ? consequent0 : (contition1 ? consequent1 : alternate)
Note: a match expression can often be a better choice than having a long
if/else if/.../else chain.
For example, in the FP adder module (modules/fp32_add_2.x), there is code like the following:
[...]
let result_fraction = if wide_exponent < u9:255 { result_fraction } else { u23:0 };
let result_exponent = if wide_exponent < u9:255 { wide_exponent as u8 } else { u8:255 };
Iterable expressions are used in counted for loops. DSLX currently supports two
types of iterable expressions, range and enumerate.
The range expression m..n represents a range of values from m to n-1.
This example will run from 0 to 4 (exclusive):
for (i, accum): (u32, u32) in u32:0..u32:4 {
There also exists a range() builtin function that performs the same operation.
enumerate iterates over the elements of an array type and produces pairs of
(index, value), similar to enumeration constructs in languages like Python or
Go.
In the example below, the loop will iterate 8 times, following the array
dimension of x. Each iteration produces a tuple with the current index (i
ranging from 0 to 7) and the value at the index (e = x[i]).
fn prefix_scan_eq(x: u32[8]) -> bits[8,3] {
let (_, _, result) =
for ((i, e), (prior, count, result)): ((u32, u32), (u32, u3, bits[8,3]))
in enumerate(x) {...
DSLX currently supports synthesis of "counted" for loops (loops that have a clear upper bound on their number of iterations). These loops are capable of being generated as unrolled pipeline stages: when generating a pipeline, the XLS compiler will unroll and specialize the iterations.
NOTE In the future support for loops with an unbounded number of iterations may be permitted, but will only be possible to synthesize as a time-multiplexed implementation, since pipelines cannot be unrolled indefinitely.
for (index, accumulator): (type-of-index, type-of-accumulator) in iterable {
body-expression
} (initial-accumulator-value)
The type annotation in the above "blueprint" is optional, but often helpful to include for increased clarity.
Because DSLX is a pure dataflow description, a for loop is an expression that produces a value. As a result, you grab the output of a for loop just like any other expression:
let final_accum = for (i, accum) in u32:0..u32:8 {
let new_accum = f(accum);
new_accum
}(init_accum);
Conceptually the for loop "evolves" the accumulator as it iterates, and ultimately pops it out as the result of its evaluation.
Add up all values from 0 to 4 (exclusive). Note that we pass the accumulator's initial value in as a parameter to this expression.
for (i, accum): (u32, u32) in u32:0..u32:4 {
accum + i
}(u32:0)
To add up values from 7 to 11 (exclusive), one would write:
let base = u32:7;
for (i, accum): (u32, u32) in u32:0..u32:4 {
accum + base + i
}(u32:0)
"Loop invariant" values (values that do not change as the loop runs) can be used
in the loop body, for example, note the use of outer_thing below:
let outer_thing: u32 = u32:42;
for (i, accum): (u32, u32) in u32:0..u32:4 {
accum + i + outer_thing
}(u32:0)
Both the index and accumulator can be of any valid type, in particular, the accumulator can be a tuple type, which is useful for evolving a bunch of values. For example, this for loop "evolves" two arrays:
for (i, (xs, ys)): (u32, (u16[3], u8[3])) in u32:0..u32:4 {
...
}((init_xs, init_ys))
Note in the above example arrays are dataflow values just like anything else. To conditionally update an array every other iteration:
let result: u4[8] = for (i, array) in u32:0..u32:8 {
// Update every other cell with the square of the index.
if i % 2 == 0 { update(array, i, i*i) } else { array }
}(u4[8]:[0, ...]);
DSLX adopts the Rust rules for semantics of numeric casts:
- Casting from larger bit-widths to smaller bit-widths will truncate (to
the LSbs).
- This means that truncating signed values does not preserve the previous value of the sign bit.
- Casting from a smaller bit-width to a larger bit-width will zero-extend if the source is unsigned, sign-extend if the source is signed.
- Casting from a bit-width to its own bit-width, between signed/unsigned, is a no-op.
#[test]
fn test_numerical_conversions() {
let s8_m2 = s8:-2;
let u8_m2 = u8:0xfe;
// Sign extension (source type is signed).
assert_eq(s32:-2, s8_m2 as s32);
assert_eq(u32:0xfffffffe, s8_m2 as u32);
assert_eq(s16:-2, s8_m2 as s16);
assert_eq(u16:0xfffe, s8_m2 as u16);
// Zero extension (source type is unsigned).
assert_eq(u32:0xfe, u8_m2 as u32);
assert_eq(s32:0xfe, u8_m2 as s32);
// Nop (bitwidth is unchanged).
assert_eq(s8:-2, s8_m2 as s8);
assert_eq(s8:-2, u8_m2 as s8);
assert_eq(u8:0xfe, u8_m2 as u8);
assert_eq(s8:-2, u8_m2 as s8);
}
Casting to an array takes bits from the MSb to the LSb; that is, the group of bits including the MSb ends up as element 0, the next group ends up as element 1, and so on.
Casting from an array to bits performs the inverse operation: element 0 becomes the MSbs of the resulting value.
All casts between arrays and bits must have the same total bit count.
fn cast_to_array(x: u6) -> u2[3] {
x as u2[3]
}
fn cast_from_array(a: u2[3]) -> u6 {
a as u6
}
fn concat_arrays(a: u2[3], b: u2[3]) -> u2[6] {
a ++ b
}
#[test]
fn test_cast_to_array() {
let a_value: u6 = u6:0b011011;
let a: u2[3] = cast_to_array(a_value);
let a_array = u2[3]:[1, 2, 3];
assert_eq(a, a_array);
// Note: converting back from array to bits gives the original value.
assert_eq(a_value, cast_from_array(a));
let b_value: u6 = u6:0b111001;
let b_array: u2[3] = u2[3]:[3, 2, 1];
let b: u2[3] = cast_to_array(b_value);
assert_eq(b, b_array);
assert_eq(b_value, cast_from_array(b));
// Concatenation of bits is analogous to concatenation of their converted
// arrays. That is:
//
// convert(concat(a, b)) == concat(convert(a), convert(b))
let concat_value: u12 = a_value ++ b_value;
let concat_array: u2[6] = concat_value as u2[6];
assert_eq(concat_array, concat_arrays(a_array, b_array));
// Show a few classic "endianness" example using 8-bit array values.
let x = u32:0xdeadbeef;
assert_eq(x as u8[4], u8[4]:[0xde, 0xad, 0xbe, 0xef]);
let y = u16:0xbeef;
assert_eq(y as u8[2], u8[2]:[0xbe, 0xef]);
}
DSLX supports Python-style bit slicing over unsigned bits types. Note that
bits are numbered 0..N starting "from the right (as you would write it on
paper)" -- least significant bit, AKA LSb -- for example, for the value
u7:0b100_0111:
Bit 6 5 4 3 2 1 0
Value 1 0 0 0 1 1 1
A slice expression [N:M] means to get from bit N (inclusive) to bit M
exclusive. The start and limit in the slice expression must be signed integral
values.
Aside: This can be confusing, because the N stands to the left of M in the
expression, but bit N would be to the 'right' of M in the classical bit
numbering. Additionally, this is not the case in the classical array
visualization, where element 0 is usually drawn on the left.
For example, the expression [0:2] would yield:
Bit 6 5 4 3 2 1 0
Value 1 0 0 0 1 1 1
^ ^ included
^ excluded
Result: 0b11
Note that, as of now, the indices for this [N:M] form must be literal numbers
(so the compiler can determine the width of the result). To perform a slice with
a non-literal-number start position, see the +: form described below.
The slicing operation also support the python style slices with offsets from
start or end. To visualize, one can think of x[ : -1] as the equivalent of
x[from the start : bitwidth - 1]. Correspondingly, x[-1 : ] can be
visualized as [ bitwidth - 1 : to the end].
For example, to get all bits, except the MSb (from the beginning, until the top element minus 1):
x[:-1]
Or to get the two most significant bits:
x[-2:]
This results in the nice property that a the original complete value can be
sliced into complementary slices such as :-2 (all but the two most significant
bits) and -2: (the two most significant bits):
#[test]
fn slice_into_two_pieces() {
let x = u5:0b11000;
let (lo, hi): (u3, u2) = (x[:-2], x[-2:]);
assert_eq(hi, u2:0b11);
assert_eq(lo, u3:0b000);
}
There is also a "width slice" form x[start +: bits[N]] - starting from a
specified bit, slice out the next N bits. This is equivalent to:
bits[N]:(x >> start). The type can be specified as either signed or unsigned;
e.g. [start +: s8] will produce an 8-bit signed value starting at start,
whereas [start +: u4] will produce a 4-bit unsigned number starting at
start.
// Identity function helper.
fn id<N: u32>(x: bits[N]) -> bits[N] { x }
#[test]
fn test_bit_slice_syntax() {
let x = u6:0b100111;
// Slice out two bits.
assert_eq(u2:0b11, x[0:2]);
assert_eq(u2:0b11, x[1:3]);
assert_eq(u2:0b01, x[2:4]);
assert_eq(u2:0b00, x[3:5]);
// Slice out three bits.
assert_eq(u3:0b111, x[0:3]);
assert_eq(u3:0b011, x[1:4]);
assert_eq(u3:0b001, x[2:5]);
assert_eq(u3:0b100, x[3:6]);
// Slice out from the end.
assert_eq(u1:0b1, x[-1:]);
assert_eq(u1:0b1, x[-1:6]);
assert_eq(u2:0b10, x[-2:]);
assert_eq(u2:0b10, x[-2:6]);
assert_eq(u3:0b100, x[-3:]);
assert_eq(u3:0b100, x[-3:6]);
assert_eq(u4:0b1001, x[-4:]);
assert_eq(u4:0b1001, x[-4:6]);
// Slice both relative to the end (MSb).
assert_eq(u2:0b01, x[-4:-2]);
assert_eq(u2:0b11, x[-6:-4]);
// Slice out from the beginning (LSb).
assert_eq(u5:0b00111, x[:-1]);
assert_eq(u4:0b0111, x[:-2]);
assert_eq(u3:0b111, x[:-3]);
assert_eq(u2:0b11, x[:-4]);
assert_eq(u1:0b1, x[:-5]);
// Slicing past the end just means we hit the end (as in Python).
assert_eq(u1:0b1, x[5:7]);
assert_eq(u1:0b1, x[-7:1]);
assert_eq(bits[0]:0, x[-7:-6]);
assert_eq(bits[0]:0, x[-6:-6]);
assert_eq(bits[0]:0, x[6:6]);
assert_eq(bits[0]:0, x[6:7]);
assert_eq(u1:1, x[-6:-5]);
// Slice of a slice.
assert_eq(u2:0b11, x[:4][1:3]);
// Slice of an invocation.
assert_eq(u2:0b01, id(x)[2:4]);
// Explicit-width slices.
assert_eq(u2:0b01, x[2+:u2]);
assert_eq(s3:0b100, x[3+:s3]);
assert_eq(u3:0b001, x[5+:u3]);
}
An infamous wrinkle is introduced for parametric functions: consider the following function:
// (Note: DSLX does not currently support the `T: type` construct shown here,
// it is for example purposes only.)
fn add_wrapper<T: type, U: type>(x: T, y: U) -> T {
x + y
}
Based on the inference rule, we know that '+' can only type check when the
operand types are the same. This means we can conclude that type T is the same
as type U. Once we determine this, we need to make sure anywhere U is used
it is consistent with the fact it is the same as T. In a sense the + operator
is "adding a constraint" that T is equivalent to U, and trying to check that
fact is valid is under the purview of type inference. The fact that the
constraint is added that T and U are the same type is referred to as
"unification", as what was previously two entities with potentially different
constraints now has a single set of constraints that comes from the union of its
operand types.
DSLX's typechecker will go through the body of parametric functions per invocation. As such, the typechecker will always have the invocation's parametric values for use in asserting type consistency against "constraints" such as derived parametric expressions, body vs. annotated return type equality, and expression inference rules.
DSLX's operator precedence matches Rust's. Listed below are DSLX's operators in descending precedence order. Binary operators at the same level share the same associativity and will be grouped accordingly.
| Operator | Associativity |
|---|---|
Unary - ! |
n/a |
as |
Left to right |
* / % |
Left to right |
+ - |
Left to right |
<< >> >>> |
Left to right |
& |
Left to right |
^ |
Left to right |
| |
Left to right |
== != < > <= >= |
Left to right |
&& |
Left to right |
|| |
Left to right |
DSLX allows specifying tests right in the implementation file via the test and
quickcheck directives.
Having key test code in the implementation file serves two purposes. It helps to ensure the code behaves as expected. Additionally it serves as 'executable' documentation, similar in spirit to Python doc strings.
Unit tests are specified by the test directive, as seen below:
#[test]
fn test_reverse() {
assert_eq(u1:1, rev(u1:1));
assert_eq(u2:0b10, rev(u2:0b01));
assert_eq(u2:0b00, rev(u2:0b00));
}
The DSLX interpreter will execute all functions that are proceeded by a test
directive. These functions should be non-parametric, take no arguments, and
should return a unit-type.
Unless otherwise specified in the implementation's build configs, functions called by unit tests are also converted to XLS IR and run through the toolchain's LLVM JIT. The resulting values from the DSLX interpreter and the LLVM JIT are compared against each other to assert equality. This is to ensure DSLX implementations are IR-convertible and that IR translation is correct.
QuickCheck is a testing framework concept founded on
property-based testing. Instead of specifying expected and test values,
QuickCheck asks for properties of the implementation that should hold true
against any input of the specified type(s). In DSLX, we use the quickcheck
directive to designate functions to be run via the toolchain's QuickCheck
framework. Here is an example that complements the unit testing of DSLX's rev
implementation from above:
// Reversing a value twice gets you the original value.
#[quickcheck]
fn prop_double_reverse(x: u32) -> bool {
x == rev(rev(x))
}
The DSLX interpreter will also execute all functions that are proceeded by a
quickcheck directive. These functions should be non-parametric and return a
bool. The framework will provide randomized input based on the types of the
arguments to the function (e.g. above, the framework will provided randomized
u32's as x).
By default, the framework will run the function against 1000 sets of randomized
inputs. This default may be changed by specifying the test_count key in the
quickcheck directive before a particular test:
#[quickcheck(test_count=50000)]
The framework also allows programmers to specify a seed to use in generating the random inputs, as opposed to letting the framework pick one. The seed chosen for production can be found in the execution log.
For determinism, the DSLX interpreter should be run with the seed flag:
./interpreter_main --seed=1234 <DSLX source file>
Footnotes
-
Otherwise there'd be a use-before-definition error. ↩