Chapel Evolution¶
Like any language, Chapel has changed over time. This page is designed to describe significant language changes that have a high likelihood of breaking existing user codes or code samples from presentations or papers that predated the changes.
Note that the compiler flag --warn-unstable
is available and can be
useful when migrating programs to the current version of the language.
The purpose of this flag is to identify portions of a program that use a
language or library feature has recently changed meaning or which is
expected to change meaning in the future.
version 1.32, September 2023¶
c_string
deprecation¶
Version 1.32 deprecates the c_string
type in user interfaces. Please
replace occurrences of c_string
with c_ptrConst(c_char)
. Note that you
need to use
or import
the CTypes
module to have access to
c_ptrConst
and c_char
types.
Here are some cases where directly replacing c_string
with
c_ptrConst(c_char)
may not work and what to do instead:
if your code is… |
update it to… |
---|---|
casting |
use a |
casting |
use a |
casting |
create a string and cast it to other type |
casting |
replace cast with |
casting |
replace cast with |
casting other type to |
create a string and call |
using |
use |
Additionally, several c_string
methods are deprecated without replacement:
.writeThis()
.serialize()
.readThis()
.indexOf()
.substring()
.size
*
An equivalent for .size
is the unstable procedure strLen(x)
in the
CTypes
module.
The default intent for arrays and records¶
In version 1.32, arrays and records now always have a default intent of
const
. This means that if arrays and records are modified inside of a
function, a coforall
, a begin
, or cobegin
, they must use a ref
intent. This also means that record methods which modify their implicit
this
argument must also use a ref
intent. Previously, the compiler would treat
these types as either const ref
intent or ref
intent, depending on if
they were modified. This change was motivated by improving the consistency
across types and making potential problems more apparent.
Since there is a lot of user code relying on modifying an outer array, the
corresponding change for forall
is still under discussion. As a result, it
will not warn by default, but modifying an outer array from a forall
might
not be allowed in the future in some or all cases.
Consider the following code segment, which contains a coforall
statement
which modifies local variables. Prior to version 1.32, this code compiled and
worked without warning.
var myInt: int;
const myDomain = {1..10};
var myArray: [myDomain] int;
coforall i in 2..9 with (ref myInt) {
myInt += i;
myArray[i] = myArray[i-1] + 1;
}
Note that to modify myInt
, an explicit ref
intent must be used whereas
myArray
can be modified freely. The changes to the default intent for
arrays is an attempt to remove this inconsistency and make the treatment of
types in Chapel more uniform. This code also modifies both myInt
and
myArray
in a way that can produce race conditions. With myInt
, it is
very apparent that there is something different than a simple serial iteration
occurring and this can signal to users to more careful inspect their code for
potential bugs. However myArray
can be used without that same restriction,
which can be a source of subtle bugs. In 1.32, the loop is written as:
var myInt: int;
const myDomain = {1..10};
var myArray: [myDomain] int;
coforall i in 2..9 with (ref myInt, ref myArray) {
myInt += i;
myArray[i] = myArray[i-1] + 1;
}
This removes the inconsistency and calls greater attention to potential race conditions.
This change also applies to procedures. Consider the following procedure:
proc computeAndPrint(ref myInt: int, myArray: []) {
...
}
It is clear that myInt
may be modified and a user of this function can save
this value beforehand if they need the value later. But without knowing what is
contained in this function, it is impossible to tell if myArray
is going to
be modified. Making the default intent for arrays const
removes this
ambiguity.
This consistency is extended to records as well. Consider the following record definition:
record myRecord {
var x: int;
proc doSomething() {
...
}
}
Without knowing what the body of doSomething
does, it is not clear
whether x
may be modified. In version 1.32, if x
is modified the
method must be marked as a modifying record using a this-intent.
record myRecord {
var x: int;
proc ref doSomething() {
...
}
}
Now it is clear that the method may modify x
.
version 1.31, June 2023¶
Version 1.31 renames and adjusts two of range’s parameters,
formerly range.boundedType
and range.stridable
,
as well as the former domain parameter domain.stridable
.
For details please see Range Types in the online documentation for Version 1.30 and Version 1.31.
Range boundedType / bounds parameter¶
Prior to Version 1.31, the boundedness of a range r
was determined
by r.boundedType
. As of 1.31, it is determined by r.bounds
.
At the same time, the type of this field changed from:
enum BoundedRangeType { bounded, boundedLow, boundedHigh, boundedNone };
to:
enum boundKind { both, low, high, neither };
This change helps make Chapel code shorter, improving its readability.
When updating your code, simply update the names accordingly. For example, from:
if myRange.boundedType == BoundedRangeType.boundedLow then ....;
to:
if myRange.bounds == boundKind.low then ....;
Range and domain stridability / strides parameter¶
Prior to Version 1.31, ranges and domains had the parameter stridable
,
which was a boolean that indicated whether the given range or domain
allowed non-unit strides.
As of 1.31, this parameter is replaced with strides
whose type is:
enum strideKind { one, negOne, positive, negative, any };
This change creates additional opportunities for optimization,
for example in the cases where the range’s stride is known at compile time
to be positive or to be -1.
This also avoids a terminology problem where stridable=false
implied,
incorrectly, that a range could not be strided. The strides
values
are now self-explanatory instead of the non-specific values
true
and false
.
When updating your code, update the field name and replace boolean values with enum values. For example:
change from… |
to… |
---|---|
|
|
|
|
|
|
|
|
another potential replacement: |
|
When getting an error like “assigning to a range with boundKind.positive from a range with boundKind.any”, insert a cast to the desired range type. Analogous updates are needed in code operating on domains.
version 1.28, September 2022¶
Version 1.28 included some significant changes to the overload resolution
rules. In addition, it enabled implicit conversion from int(t)
to
uint(t)
. This section discusses some example programs that behave
differently due to these changes.
See also:
Behavior Differences for Mixes of Signed and Unsigned¶
Prior to 1.28, numeric operations applied to a mix of signed and unsigned types could have surprising results by moving the computation from a particular bit width to another—or by moving it from an integral computation to a floating point one.
For example:
var myInt:int = 1;
var myUint:uint = 2;
var myIntPlusUint = myInt + myUint; // what is the type of `myIntPlusUint`?
Before 1.28, this program would result in compilation error, due to an
error overload of operator +
in the standard library.
Version 1.28 adds the ability for an int
to implicitly convert to
uint
and removes the error overload. As a result, the uint
version of operator +
is chosen, which results in myIntPlusUint
having type uint
.
This behavior can also extend to user-defined functions. Consider a function
plus
defined for int
, uint
, and real
:
proc plus(a: int, b: int) { return a + b; }
proc plus(a: uint, b: uint) { return a + b; }
proc plus(a: real, b: real) { return a + b; }
var myInt:int = 1;
var myUint:uint = 2;
var myIntPlusUint = plus(myInt, myUint);
In 1.27 the call to plus
would resolve to the real
version because
int
could not implicitly convert to uint
, but both int
and uint
could implicitly convert to real(64)
. As a result, myIntPlusUint
had
the type real
. This change from integral types to floating point types
could be very surprising.
In 1.28 the call to plus
resolves to the uint
version, and
myIntPlusUint
has type uint
.
This behavior also applies to int
and uint
types with smaller widths:
var myInt32:int(32) = 1;
var myUint32:uint(32) = 2;
var myInt32PlusUint32 = myInt32 + myUint32;
In 1.27, the int(64)
+
operator is chosen (because both int(32)
and
uint(32)
can implicitly convert to int(64)
), which results in
myInt32PlusUint32
having type int(64)
. This could be surprising when
explicitly working with 32-bit numbers.
In contrast, in 1.28, due to the ability for int(32)
to implicitly
convert to uint(32)
, the uint(32)
version of the +
operator
is chosen and myInt32PlusUint32
has type uint(32)
.
Param Expression Behavior¶
Some expressions consisting of mixed-type literal or param
values now
have different behavior. For example:
var x = 1:int(8) + 2; // what is the type of `x` ?
Note in this example that the literal 2
is a param
with type
int(64)
and that 1:int(8)
is a param
with type int(8)
.
In 1.27, this program would output int(8)
, because the overload
resolution rules would favor the +
overload using the type of the
non-default-sized param
. The result is that in 1.27, x
had type
int(8)
.
In 1.28, the rules are simpler and a closer match to the corresponding
case with regular variables (myInt8 + myInt64
). There is no longer
any special behavior for non-default-sized param
. As a result, the
value x
now has type int(64)
.
For similar reasons, the type of nI
in the following code is now
int(64)
where previously it was int(32)
:
const nI = ((-2):int(32))**53;
A similar change can also appear with range literals that use mixed type
param
lower and upper bounds. The following range construction also
makes use of the new implicit conversion from int(t)
to uint(t))
:
var r8 = 1:int(8)..100:uint(8);
writeln(r8.type:string);
In 1.27, this would generate a range with index type int(16)
. In
1.28, it produces a range with index type uint(8)
.
Speaking of range literal construction, a range like 1:int(8)..10
still produces an int(8)
range in 1.28. However, as we have
discussed, something like 1:int(8) + 10
would result in an
int(64)
. For now, the range implementation has been adjusted to
preserve the old behavior specifically for the ..
operator. However,
this may change in a future release.
Change for some mixed int/uint overloads¶
This example shows a change in behavior for two overloads where one is
int
and the other is uint
:
proc dbm(a:int(8)) { writeln("dbm int8"); }
proc dbm(a:uint(64)) { writeln("dbm uint64"); }
dbm(42:int(64));
Previous to 1.28, this program would call the int(8)
version of the
function. It can do that because the compiler knows that the param
value 42
will fit into an int(8)
. Such a conversion is called a
param
narrowing conversion. However, in 1.28, this function now calls
the uint(64)
version of the function. The main reason for this is
that the 1.28 rules prefer to not do param
narrowing conversion when
another candidate does not need it. In this case, int
to uint
is
not a param
narrowing conversion so that is preferred.
Change for function visibility / shadowing¶
The new overload resolution rules in 1.28 consider function visibility or shadowing before considering how well the arguments match. Consider this example:
proc f(arg: int) { writeln("f int"); }
proc main() {
proc f(arg) { writeln("f generic"); }
f(1); // which `f` does this call?
}
Inside of proc main
, the call to f
now resolves to the generic
inner function. In contrast, in version 1.27, the outer
proc f(arg: int)
would be called.
version 1.22, April 2020¶
0- vs. 1-based Indexing¶
Version 1.22 makes a major breaking change to Chapel with respect to indexing for cases that involve implicit indices. Historically, Chapel has used 1-based indexing for such cases, where it now uses 0-based indexing.
The major types that are affected by this change are tuples, strings,
bytes
, and lists. In addition, arrays that don’t have a
well-defined index set also start at 0. Such cases include array
literals or inferred-type arrays formed by capturing a general
iterator expression.
This change also has a ripple-down effect to features and routines
related to these types. For example, varargs arguments are equivalent
to tuples in Chapel, so inherit their 0-based indexing. Similarly,
queries on rectangular domains and arrays are based on tuples,
so their dimensions are now numbered from 0 as well.
Certain library routines such as find()
on strings used to return 0
when no match was found, but now return -1 in order to avoid returning
a legal string index.
The following sections summarize the rationale for this change and then provide some tips for updating existing Chapel code.
Rationale for 0- vs. 1-based Indexing¶
In the original design of Chapel, we hoped to make the language as neutral to 1- vs. 0-based indexing as possible, to avoid running afoul of the strong emotions that such choices evoke in users when it doesn’t match their preference. As a result, Chapel’s primary types for parallel computation on regular collections of data—namely, its ranges and rectangular domains, as well as rectangular arrays defined by ranges or domains—require users to specify both low and high bounds. Happily, these core features are not affected by this change in Chapel 1.22, so codes relying solely on such features will not require updates.
However, for other types such as tuples and strings, we were forced to make a decision. At the time of Chapel’s inception, the main languages from which we were trying to attract users were C/C++, Java, Fortran, and Matlab. Since half of these languages used 0-based indexing and the other half used 1-based, there didn’t seem to be an obvious best answer. In the end, we decided to go with 1-based indexing on the argument that we were striving to create a productive language, and that counting from 1 is arguably most natural for most people.
Over time, however, the vast majority of newer languages that we look to for users or inspiration—most notably Python, Swift, and Rust—have been almost exclusively 0-based. Meanwhile, very few notable new languages have used 1-based indexing.
Furthermore, when polled, the vast majority of active Chapel users expressed a strong preference for 0-based programming, given the choice (though there were also notable outliers, particularly from the Fortran community). We also realized (a) that Chapel’s design should be more concerned with lowering barriers for existing programmers than for non-programmers; and (b) that even though we had arguably biased the original design in favor of Fortran programmers, most of Chapel’s early adopters have come from C/C++ and Python backgrounds.
Based on this, we undertook an experiment to see what it would take to convert from 1-based to 0-based programming. Reviewing Chapel’s ~10,000 tests and modules resulted in changes to ~1,000 of them. We also updated some significant applications such as Arkouda and Cray HPO. While the overall effort of making the change was not insignificant, it also wasn’t particularly difficult for the most part. Overall, our finding was that in cases where the changes weren’t simply neutral in their impact on style, it almost always benefitted the code in terms of clarity, because there tended to be fewer adjustments of +/- 1 in the code.
For these reasons, we decided to bite the bullet and make the switch now, while we felt we still could, rather than later when it would clearly be too late to do so and cause more of a revolt among our users.
Index-neutral Features¶
This experience also led to a number of new programming features in
Chapel 1.21 designed to help write code in more of an index-neutral
style. Chief among these are new .indices
queries on most of the
relevant types as well as support for loops over heterogeneous tuples.
We also introduced features that we found to be useful in updating
code, such as support for open-interval ranges and .first
and
.last
queries on enumerated types. To this end, even though Chapel
still has cases that require making this 0- vs. 1-based indexing
decision, we encourage code to be written in an index-neutral style
whenever possible, and believe that most common code patterns can be.
Tips for Updating Existing Chapel code¶
The following are some tips for updating codes based on our experiences:
First, updating code is easiest when it has some sort of testing infrastructure that can be used to validate that its behavior is unchanged. If you don’t already have such testing for your code, it may be worthwhile to invest in creating some before attempting this upgrade.
Next, when transitioning code to Chapel 1.22, make sure to compile it with neither
--fast
nor--no-checks
enabled so that bounds checks are turned on in the generated code. In cases where a program is accessing all of the elements of a collection (as is common for tuples) this will help identify data structures that require updates. When you do get an out-of-bounds error, don’t simply update the specific access, but use it as a cue to look through the code for other references to that variable that will also need updating.When possible, try rewriting your updated code to use an index-neutral style of programming. For example, given code like this:
var t: 2*int = ...; var x = t(1), y = t(2); for i in 1..2 do writeln("t(", i, ") = ", t(i));
It would be reasonable to rewrite it like this:
var t: 2*int = ...; var x = t(0), y = t(1); for i in 0..1 do writeln("t(", i, ") = ", t(i));
But arguably preferable to update it like this:
var t: 2*int = ...; var (x, y) = t; for i in t.indices do writeln("t(", i, ") = ", t(i));
If you have a pattern that you’re trying to write in an index-neutral style, but can’t, don’t hesitate to ask for tips.
Some common pitfalls to check for in your code include:
Search for queries on the dimensions of rectangular domains and arrays. For example,
myDomain.dim(1)
,myDomain.low(1)
,myDomain.high(1)
, ormyDomain.stride(1)
will need to be updated to reflect that array dimensions now count from 0 rather than 1. These will result in out-of-bounds errors in cases where you query all dimensions of an array, making them easy to find; but it can be worthwhile to grep your code for such patterns to make sure you don’t miss any.Also search for instances of
find()
orrfind()
that are relying on comparisons to zero/nonzero values, and update them to compare against -1. For example, patterns likeif mystring.find('z')
need to be updated toif mystring.find('z') != -1
.Search for instances of
split()
. A common idiom is to writevar substrs = mystring.split(5);
and then to index into the result usingsubstrs[1]
,substrs[2]
, etc. Since this is an instance of capturing an iterator expression, you’ll either need to subtract one from the indices, or else declare substrs to have a specific type, likevar substrs: [1..5] string = mystring.split(5);
Search for varargs functions and make sure they are updated to use 0-based indexing or index-neutral features.
Search for any calls to
Reflection.getField*()
and update those the cases that use integer indices to reflect 0-based numbering.Look for any calls on lists that use explicit offsets, as these will likely need updates. For example
mylist.pop(1);
will need to becomemylist.pop(0);
Some other common string patterns to look for in your code that may indicate something requiring an update include:
1..
[1]
(1)
[2]
(2)
Think about whether there are other places in your code that compute index values numerically yet which don’t have obvious syntactic cues.
Need Help?¶
If you are able to share your code with us and would like help updating it to Chapel 1.22, please don’t hesitate to ask for help. Given our experience in updating the Chapel code base itself, we have found it fairly easy to update most codes, even when we’re unfamiliar with them.
version 1.21, April 2020¶
Version 1.21 made several improvements related to record initialization, assignment, and deinitialization.
In summary:
Some patterns of default initialization followed by assignment are now converted to initialization. See split initialization.
Some patterns of copy initialization followed by deinitialization are converted to move initialization. See copy elision.
The result of a nested call expression can now be deinitialized at the end of the containing statement. See deinitialization point of nested call expressions.
split initialization¶
Split initialization a new language feature in 1.21 that is described in the language specification - see Split Initialization.
Consider the following example:
var x: myRecord; // default-initialization in 1.20
x = new myRecord(); // assignment in 1.20 -- initialization in 1.21
In 1.21, instead of default-initializing x
and then assigning to it,
x
will be initialized on the second line.
Note that split initialization also changes the copy and assignment
behavior of out
intent formal arguments.
Occasionally programs that are written to test assignment (separately from copy initialization) need to avoid split initialization. One way to do so is to add a mention of the variable immediately after it is declared, as in the following code:
var x: myRecord;
x; // adding this mention prevents split-initialization
// instead, x is default-initialized at its declaration point above
x = new myRecord();
copy elision¶
Copy elision a new language feature in 1.21. When the last mention of a variable is the source of a copy-initialization, the copy-initialization is replaced by move-initialization.
For example:
class MyClass {
var field;
proc init(in arg) {
this.field = arg;
}
}
proc copyElisionExample() {
var a = new myRecord();
var b = a; // now move-initializes `b` from `a`
return new MyClass(b); // now move-initializes the field from `b`
}
deinitialization point of nested call expressions¶
In 1.20, all variables are deinitialized at the end of the enclosing block. That changed in 1.21. Compiler-introduced temporary variables storing the result of a nested call expression can now be deinitialized at the end of a statement. In particular, results of nested call expressions are now deinitialized at the end of the statement unless the statement is initializing a user variable.
For example:
proc makeRecord() {
return new myRecord();
}
proc f(arg) {
return arg;
}
proc deinitExample() {
f(makeRecord());
// Compiler converts the above statement into
// var tmp = makeRecord();
// f(tmp);
// In 1.20, tmp is destroyed at the end of the block.
// In 1.21, tmp is destroyed at the end of the above statement.
var x = f(makeRecord());
// In both 1.20 and 1.21, the temporary storing the result of
// `makeRecord()` is deinitialized at the end of the block.
}
version 1.20, September 2019¶
Version 1.20 made language changes that address problems with classes.
In summary:
variables of class type can no longer store nil by default but can opt-in to possibly being nil with ?. See nilability changes
certain casts have changed behavior to support nilability changes See nilability and casts
un-decorated class types such as MyClass (as opposed to borrowed MyClass) now have generic management See undecorated classes have generic management
arguments with owned or shared declared type now use const ref default intent rather than in intent. See new default intent for owned and shared
new C
now creates an owned C rather than a borrowed C See new C is owned
nilability changes¶
Previous to 1.20, variables of class type could always store nil
. In
1.20, only nilable class types can store nil
. Non-nilable class types
and nilable class types are different types. A class type expression
such as borrowed C
indicates a non-nilable class type.
As an aid in migrating code to this change, the flag --legacy-classes
will disable this new behavior.
Consider the following example:
class C {
var x:int;
}
var a: borrowed C = (new owned C()).borrow();
In 1.19, variables of type borrowed C
could store nil
:
var b: borrowed C = nil;
var c: borrowed C;
a = nil;
The 1.20 compiler will report errors for all 3 of these lines. To resolve
the errors, it is necessary to use a nilable class type. Nilable class
types are written with ?
at the end of the type. In this example:
var a: borrowed C? = (new owned C()).borrow();
var b: borrowed C? = nil;
var c: borrowed C?;
a = nil;
Implicit conversions are allowed from non-nilable class types to nilable class types.
When converting variables to nilable types to migrate code, there will be
situations in which it is known by the developer that a variable cannot
be nil
at a particular point in the code. For example:
proc f(arg: borrowed C) { }
proc C.method() { }
config const choice = true;
var a: owned C?;
if choice then
a = new owned C(1);
else
a = new owned C(2);
f(a);
a.method();
Errors on the last two lines can be resolved by writing
f(a!);
a!.method();
where here the !
asserts that the value is not nil
and it can
halt if the value is nil
.
Note that in prototype
and implicit file-level modules, the compiler
will automatically add !
on method calls with nilable receivers
(i.e. in the a.method()
case above).
In the above case, a cleaner way to write the conditional would be to create a function that always returns a value or throws if there is a problem. For example:
proc makeC() throws {
var a: owned C?;
if choice then
a = new owned C(1);
else
a = new owned C(2);
return a:owned C; // this cast throws if a stores nil
}
proc main() throws {
var a:owned C = makeC();
f(a);
a.method();
}
nilability and casts¶
Because casts to class types should necessarily return something of the
requested type, and because many class types now cannot store nil
,
certain patterns involving casts will need to change to work with 1.20.
class downcasts¶
In a class downcast, a class is casted to a subtype. If the dynamic type
of the variable does not match the requested subtype, the downcast fails.
In 1.19, a failed downcast would result in nil
. In 1.20, a failed
downcast will result in nil
only if the target type is nilable and
will throw an error otherwise.
For example:
class Parent { }
class Child : Parent { }
var p:borrowed Parent = (new owned Parent()).borrow();
var c:borrowed Parent = (new owned Child()).borrow();
writeln(c:Child?); // downcast succeeds
writeln(c:Child); // downcast succeeds
writeln(p:Child?); // this downcast fails and results in `nil`
writeln(p:Child); // this downcast fails and will throw a ClassCastError
casting C pointers to classes¶
Casts from c_void_ptr
to class types were previously allowed. However,
since c_void_ptr
can store NULL
, this case needs adjustment
following the nilability changes. Additionally, since c_void_ptr
refers to a C pointer, and C pointers are manually managed (i.e. you call
free
on them at the appropriate time), it makes the most sense
for casts from c_void_ptr
to end up with an unmanaged type.
Consider the following example:
class C {
var x:int;
}
var myC = new owned C();
var ptr:c_void_ptr = myC.borrow(); // store the instance in a C ptr
Now we can cast from ptr
to the class type:
var c = ptr:C; // cast from a C pointer to the borrowed type
This example would work in 1.19. In 1.20, it needs to be updated to
cast to unmanaged C?
:
var c = ptr:unmanaged C?;
As with other values of type unmanaged C?
, from there it can:
be borrowed, e.g.
c.borrow()
have
!
applied to convert to a non-nilable value or halt, e.g.c!
be cast to a non-nilable type, throwing if it is
nil
, e.g.c:borrowed C
undecorated classes have generic management¶
Undecorated classes now have generic management. As an aid in migrating
code to this change, the flag --legacy-classes
will disable this
new behavior.
Supposing that we have a class C
declaration as in the following:
class C {
var x:int;
}
Code using C
might refer to the type C
on its own or it might use
a decorator to specify memory management strategy, as in borrowed C
.
The type expression C
was the same as borrowed C
in 1.18 and
1.19 but now means generic management. For example, in the following code:
var myC:C = new owned C();
myC
previously had type borrowed C
, and was initialized using
including an implicit conversion from owned C
to borrowed C
. In 1.20,
myC
has type owned C
. Since the variable’s type expression is
generic management, it takes its management from the initializing
expression.
This change combines with the nilability changes described above to prevent compilation of existing code like the following:
var x:C;
Knowing that C
now cannot store nil
, one might try to update this
program to:
var x:C?;
However this does not work either. C?
indicates a nilable class type
with generic management, and a variable with generic type cannot be
default-initialized.
To update such a variable declaration to 1.20, it is necessary to include
a memory management decorator as well as ?
. For example:
var x:borrowed C?;
The resulting variable will initially store nil
.
new C is owned¶
Supposing that C is a class type, new C() was equivalent to new borrowed C() before this release - meaning that it resulted in something of type borrowed C. However, it is now equivalent to new owned C() which produces something of type owned C.
version 1.18, September 2018¶
Version 1.18 includes many language changes that address problems with classes.
In summary:
constructors are deprecated and replaced with initializers See initializers replace constructors
memory management for class types has changed See class memory management
override is now required on overriding methods See overriding methods must be marked
initializers replace constructors¶
Code that contained user-defined constructors will need to be updated to use an initializer. For example:
record Point {
var x, y: real;
proc Point() {
x = 0;
y = 0;
writeln("In Point()");
}
proc Point(x: real, y: real) {
this.x = x;
this.y = y;
writeln("In Point(x,y)");
}
}
var a:Point;
var b = new Point(1.0, 2.0);
will now compile with deprecation warnings. Here is the same program updated to use initializers:
record Point {
var x, y: real;
proc init() {
x = 0;
y = 0;
writeln("In Point.init()");
}
proc init(x: real, y: real) {
this.x = x;
this.y = y;
writeln("In Point.init(x,y)");
}
}
var a:Point;
var b = new Point(1.0, 2.0);
The change to initializers is much more than a change in the name of the method. See the language specification for further details.
class memory management¶
Before 1.18, if C
is a class type, a variable of type C
needed
to be deleted in order to prevent a memory leak. For example:
class C {
var x: int;
}
proc main() {
var instance: C = new C(1);
delete instance;
}
Version 1.18 introduced four memory management strategies that form part of a class type and are used with new expressions:
owned C
owned
classes will be deleted automatically when theowned
variable goes out of scope, but only oneowned
variable can refer to the instance at a time. Such instances can be created withnew owned C()
.shared C
shared
classes will be deleted when all of theshared
variables referring to the instance go out of scope. Such instances can be created withnew shared C()
.borrowed C
refers to a class instance that has a lifetime managed by another variable. Values of type
borrowed C
can be created withnew borrowed C()
, by coercion from the other classC
types, or by explicitly calling the.borrow()
method on one of the other classC
types.new borrowed C()
creates a temporary instance that will automatically be deleted at the end of the current block.unmanaged C
the instance must have delete called on it explicitly to reclaim its memory. Such instances can be created with
new unmanaged C()
.
Further note that the default is borrowed
, that is:
C
is now the same as
borrowed C
new C()
is now the same as
new borrowed C()
Now, back to the example above. There are several ways to translate this program.
First, the most semantically similar option is to replace uses of C
with unmanaged C
:
class C {
var x: int;
}
proc main() {
var instance: unmanaged C = new unmanaged C(1);
delete instance;
}
Using unmanaged
allows a Chapel programmer to opt in to manually
managing the memory of the instances.
A reasonable alternative would be to translate the program to use
owned C
:
class C {
var x: int;
}
proc main() {
var instance: owned C = new owned C(1);
// instance will now be automatically deleted at the end of this block
}
If the program does not explicitly use owned C
, it can rely on
new C()
being equivalent to new borrowed C()
:
class C {
var x: int;
}
proc main() {
var instance: C = new C(1);
// instance will now be automatically deleted at the end of this block
}
See the Class New section in the Classes chapter of the language specification for more details.
overriding methods must be marked¶
Before 1.18, a class inheriting from another class can create an overriding method that is a candidate for virtual dispatch:
class Person {
var name: string;
proc greet() {
writeln("Hello ", name, "!");
}
}
class Student: Person {
var grade: int;
proc greet() {
writeln("Hello ", name, ", welcome to grade ", grade);
}
}
proc main() {
var person: Person = new Student("Jeannie", 5);
person.greet(); // uses the run-time type of person (Student)
// and virtually dispatches to Student.greet()
}
Now such overriding methods must be marked with the override keyword:
class Person {
var name: string;
proc greet() {
writeln("Hello ", name, "!");
}
}
class Student: Person {
var grade: int;
override proc greet() {
writeln("Hello ", name, ", welcome to grade ", grade);
}
}
proc main() {
var person: Person = new Student("Jeannie", 5);
person.greet(); // uses the run-time type of person (Student)
// and virtually dispatches to Student.greet()
}
version 1.15, April 2017¶
Version 1.15 includes several language changes to improve array semantics.
In summary:
arrays are always destroyed when they go out of scope and in particular will not be preserved by use in begin. See array lexical scoping.
the array alias operator => has been deprecated in favor of creating references to an array or a slice of an array with ref or const ref. See array alias operator deprecated.
arrays now return by value by default instead of by ref. See arrays return by value by default.
arrays now pass by ref or const ref by default, depending on whether or not the formal argument is modified. See array default intent.
Additionally, the default intent for record method receivers has changed:
the method receiver for records is passed by ref or const ref by default, depending on whether or not the formal argument is modified. See record this default intent.
array lexical scoping¶
As described in the language changes for 1.12 in lexical scoping, using arrays beyond their scope is a user error. While such a program was in error starting with Chapel 1.12, such a pattern worked until Chapel 1.15.
For example, this program will probably crash in Chapel 1.15:
proc badBegin() {
var A: [1..10000] int;
begin {
A += 1;
}
// Error: A destroyed here at function end, but the begin could still
// be using it!
}
Similarly, using a slice after an array has been destroyed is an error:
proc badBeginSlice() {
var A: [1..10000] int;
// slice1 sets up a slice using the => operator
// note that the => operator is deprecated (see below)
var slice1 => A[1..1000];
// slice2 sets up a slice by creating a reference to it
ref slice2 = A[1..1000];
// either way, using the slice in a begin that can continue
// after the function declaring the array exits is an error
begin {
slice1 += 1;
slice2 += 1;
}
// Error: A destroyed here at function end, but the begin tries to
// use it through the slices!
}
array alias operator deprecated¶
The array alias operator, =>, has been deprecated in Chapel 1.15. Previously, the supported way to declare one array that aliases another (or a slice of another) was to use =>. Now, the supported way to do that is to use a ref or const ref variable:
For example, before Chapel 1.15 you might have written:
// pre-1.15
var A:[1..10] int;
// set up a const alias of A
const alias => A;
// set up a mutable slice of A
var slice => A[2..5];
// set up a re-indexing slice of A
var reindex:[0..9] => A;
In Chapel 1.15, use ref or const ref to create the same pattern:
var A:[1..10] int;
// set up a const alias of A
const ref alias = A;
// set up a mutable slice of A
ref slice = A[2..5];
// set up a re-indexing slice of A
ref reindex = A.reindex({0..9});
arrays return by value by default¶
Before Chapel 1.15, returning an array would return the array by reference. Now arrays return by value by default. That is, the act of returning an array can make a copy:
var A: [1..4] int;
proc returnsArray() {
return A;
}
ref B = returnsArray();
B = 1;
writeln(A);
// outputs 1 1 1 1 historically
// outputs 0 0 0 0 after Chapel 1.15
This behavior applies to array slices as well.
The old behavior is available with the ref return intent. Note though that returning a ref to a local array is an error just like it is an error to return a local int variable by ref.
proc returnsArrayReference() ref {
return A;
}
array default intent¶
Before 1.15, the default intent for arrays was ref. The rationale for this feature was that it was a convenience for programmers who are used to modifying array formal arguments in their functions. Unfortunately, it interacted poorly with return intent overloading. Additionally, the implementation had several bugs in this area.
The following example shows how it might be surprising that return intent overloading behaves very differently for arrays than for other types. As the example shows, this issue affects program behavior and not just const-checking error messages from the compiler.
// First, let's try some of these things with an
// associative array of ints:
{
var D:domain(int);
var A:[D] int;
// This adds index 1 to the domain, implicitly
A[1] = 10;
writeln(D.member(1)); // outputs `true`
// This will halt, because index 2 is not in the domain
//var tmp = A[2];
// This will also halt, for the same reason
//writeln(A[3]);
}
// Now, let's try the same things with an array of arrays:
{
var D:domain(int);
var AA:[D] [1..4] int;
var value:[1..4] int = [10,20,30,40];
// This adds index 4 to the domain, implicitly
AA[4] = value;
writeln(D.member(4)); // outputs `true`
// This will halt, because index 5 is not in the domain
//var tmp = AA[5];
// It seems that this *should* halt, but it does not (pre 1.15)
// Instead, it adds index 6 to the domain
writeln(AA[6]);
writeln(D.member(6)); // outputs `true` !
}
See GitHub issue #5217 for more examples and discussion.
In order to make such programs less surprising, version 1.15 changes the default intent for arrays to ref if the formal argument is modified in the function and const ref if not. As a result, the above example behaves similarly for an associative array of integers and an associative array of dense arrays.
For example, in the following program, the default intent for the formal argument x is ref:
proc setElementOne(x) {
// x is modified, so x has ref intent
x[1] = 1;
}
var A:[1..10] int;
setElementOne(A);
In contrast, in the following program, the default intent for the formal argument y is const ref:
proc getElementOne(y) {
// y is not modified, so y has const ref intent
var tmp = y[1];
}
const B:[1..10] int;
getElementOne(B);
record this default intent¶
Before 1.15, the default intent for the implicit this argument for record methods was implemented as ref but specified as const ref. In 1.15, this changed to ref if the formal this argument is modified in the body of the function and const ref if not.
See GitHub issue #5266 for more details and discussion.
record R {
var field: int;
proc setFieldToOne() {
// this is modified, so this-intent is ref
this.field = 1;
}
proc printField() {
// this is not modified, so this-intent is const ref
writeln(this.field);
}
}
version 1.13, April 2016¶
ref return intent¶
Previous versions of Chapel included an implicit setter param of type bool for ref return intent functions. In addition, the compiler created a getter and setter version of each ref return intent function. The getter version would return an rvalue, and the setter version would return an lvalue by ref. For example:
var x = 1;
proc refToX() ref {
if setter then
return x; // setter version
else
return 0; // getter version
}
refToX() = 3; // uses the setter version
writeln(x); // prints 3
var tmp = refToX(); // uses the getter version
writeln(tmp); // prints 0
This functionality has changed with version 1.13. It is still possible to write a getter and a setter, but these must be written as pair of related functions:
var x = 1;
// setter version
proc refToX() ref {
return x;
}
// getter version
proc refToX() {
return 0;
}
refToX() = 3; // uses the setter version
writeln(x); // prints 3
var tmp = refToX(); // uses the getter version
writeln(tmp); // prints 0
In some cases, when migrating code over to the new functionality, it is useful to put the old ref return intent function into a helper function with an explicit param setter argument, and then to call that function from the getter or setter.
version 1.12, October 2015¶
lexical scoping¶
Prior to version 1.12 of Chapel, variables could be kept alive past their lexical scopes. For example:
{
var A: [1..n] real;
var count$: sync int;
var x: real;
begin with (ref x) {
... A ...;
... count$ ...;
... x ...;
}
// ^^^ this task and its references to A, count$, and x could outlive
// the scope in which those variables are declared.
} // So, previously, Chapel kept these variables alive past their
// logical scope.
Disadvantages of this approach included:
It moves logical stack variables (like x and count$ above) to the heap.
It complicates memory management by incurring reference counting overhead—or causing memory leaks in cases where reference counting hadn’t been added.
It was not particularly valued or utilized by users.
It was arguably surprising (“x still exists even though it left scope?”).
As of Chapel 1.12 (and moreso in subsequent releases), the implementation no longer provides this property. Instead, it is a user error to refer to a variable after it has left scope. For example:
var flag$: sync bool; // flag$ starts empty
{
var x: real;
begin with(ref x) { // create task referring to x
flag$; // block task until flag$ is full
... x ... // user error: access to x occurs after it leaves scope
} // end task
} // x`s scope ends
flag$ = true; // fill flag$ only after x's scope closes
Code that refers to lexically scoped variables within tasks in this manner should use sync variables or blocks in order to guarantee the tasks’s completion before the enclosing block exits. Note that the more commonly used cobegin, coforall, and forall statements already guarantee that the tasks they create will complete before the enclosing block exits.
version 1.11, April 2015¶
forall intents¶
In previous versions of Chapel, the bodies of forall-loops have referred to all lexically visible variables by reference. In this release of Chapel, such variables are treated more consistently with the task intent semantics and syntax introduced in versions 1.8 and 1.10 respectively (described below).
Specifically, prior to this release, a loop like the following would represent a data race:
var sum = 0.0;
forall a in A do sum += a;
since multiple iterations of the loop could execute simultaneously, read
the identical value from the shared variable sum
, update it, and write the
result back in a way that could overwrite other simultaneous updates.
Under the new forall intent semantics, such variables are treated as though
they are passed by “blank intent” to the loop body (so const
for variables of
scalar type like sum
, preventing races in such cases). This mirrors the task
intent semantics for variables referenced within begin, cobegin
, and coforall
constructs. As in those cases, a user can specify semantics other than the
default via a with-clause. For example, to restore the previous race-y
semantics, one could write:
var sum = 0.0;
forall a in A with (ref sum) do
sum += a;
(Of course, the safe way to write such an idiom would be to use a
reduction, or a synchronization type like sync
or atomic
).
type select statement¶
Chapel has traditionally supported a type select
statement that was
like a select
statement for types. However, this seemed inconsistent with the
fact that other constructs like if...then
operate on types directly. For that
reason, this release removed support for type select x
. Instead, use the
equivalent select x.type
.
version 1.10, October 2014¶
task intents syntax¶
Task intent clauses were added to Chapel in version 1.8 to support passing variables by reference into tasks. Since then, the need to pass variables by other intents and into other parallel constructs has arisen. But, the previous syntax was problematic to extend to other intents, while also generating syntactic ambiguities for other additions we wished to make to the language.
For these reasons, a new task intent syntax was designed to cleanly support
intents other than ref
(particularly in looping contexts), to address the
pending ambiguity, and to better reflect the similarity of task intents to
formal argument lists. Where previously, task constructs could be followed by a
ref
clause, they can now be followed by a with
clause that takes a list of
intents and variables, specifying how to pass them into the task.
Thus, where one would have previously written:
begin ref(x) update(x);
cobegin ref(x, y) {
process(x);
process(y);
}
coforall x in A ref(y) {
process(x, y);
}
you would now write:
begin with (ref x) update(x);
cobegin with(ref x, ref y) {
process(x);
process(y);
}
coforall x in A with (ref y) {
process(x, y);
}
As of the 1.10 release, only ref
intents are supported, though we plan to
expand this set of intents for the 1.11 release while also extending
forall-loops to support task intents.
‘var’ function return intents changed to ‘ref’¶
A var
function return intent has traditionally been used to indicate that
a call to the function (referred to as a var function) could appear in either
an r-value or l-value context. The var
keyword was chosen since the function
could be used in the same contexts as a variable could.
Since that time, the ref
keyword has been introduced into Chapel to
support passing variables by reference to functions. Since returning an
expression by reference supports similar capabilities as var
functions require,
while also being less unusual/more orthogonal, this release replaces var
function return intents with ref
intents.
Thus, where one would previously write:
proc getFoo() var { ... }
now you would write:
proc getFoo() ref { ... }
The var
as a return intent is deprecated and generates a warning for the
current release, after which it will be removed.
version 1.9, April 2014¶
operator precedence changes to benefit common cases¶
Historically, Chapel’s operator precedence choices have tended to follow the lead of C for operators that are common to both languages, figuring that following an established convention would be better than forging our own path.
With this change, we modified the precedence of bitwise operators to better
reflect what we think it intuitive to users and correct what is viewed in many
circles to be a regrettable mistake in C. At the same time, we changed the
binding of in
and ..
to support some other Chapel idioms more naturally,
like 1..10 == 1..10
. To see the current operator precedence, refer to the
Quick Reference sheet.
improved interpretation of {D}¶
Historically, for a domain D, Chapel has interpreted {D}
as being equivalent to
D
, inheriting a precedent of sorts set by the ZPL language, and dating from a
time when we used square brackets for both domain literals and array types.
With this change, we began interpreting {D}
as a domain literal with a
single index, D
(i.e., an associative domain of domains). Associative domains
of domains are not yet implemented in the language, so the new syntax is not
yet useful, but at least the incongruity of ignoring the curly brackets has
been removed.
version 1.8, October 2013¶
task functions and intents; ref-clauses Chapel has three constructs for
creating tasks: begin
, cobegin
, and coforall
. Historically, variable references
within tasks followed standard lexical scoping rules. For example, the
following code:
var x = 0;
begin writeln(x);
x += 1;
could print either the value 0 or 1, depending on whether the writeln()
task was executed before or after the increment of x
.
With this change, we view the creation of a task as an invocation of a task function
— a compiler-created function that implements the task. Any references
to variables outside of the task’s scope (like x
in the example above) are
treated as implicit arguments to the task function, passed by blank intent.
Thus, when x
is an integer, as in the above code, the task will always
print the value of 0, even if the increment of x
is executed before the
writeln()
task, since the value of x
will have been passed to the task function
by blank intent (implying a const
copy for integer arguments). In contrast, if
x were a sync variable in the example above, the blank intent would cause it to
be passed by reference to the task, permitting the task to see either of the
values 0 or 1.
To return to the previous behavior, a ref-clause can be added to the tasking construct to indicate that a variable should be passed to the task function by reference rather than blank intent. For example, the following code:
var x = 0;
begin ref(x) writeln(x);
x += 1;
would revert to the previous behavior, even if x
were an integer.
For more information on this feature, please refer to the Task Intents section of the Task Parallelism and Synchronization chapter of the language specification.
version 1.6, October 2012¶
domain literals¶
Chapel’s domain literals were historically specified using
square brackets, based on ZPL’s region syntax. Thus [1..m, 1..n]
represented an
m × n index set.
In this change, we made domain literals use curly brackets in order to
reflect their role as sets of indices, and also to make square brackets
available for supporting array literals. Thus, {1..m, 1..n}
is an m × n index
set, [1.2, 3.4, 5.6]
is a 3-element array of reals and [1..m, 1..n]
is a
2-element array of ranges.
Emacs users working on updating existing code can use the following recipe to update old-style domain literals to the new syntax:
M-x query-replace-regexp: \([=|,] *\)\[\(.*?\)\]\([;|)]\)
with: \1{\2}\3
zippered iteration¶
Zippered iteration in Chapel was traditionally supported
simply by iterating over a tuple of values. For example, forall (i,a)
in
(1..n, A)
would iterate over the range 1..n
and the n-element array
A
in a zippered manner.
In this change, we introduced the zip keyword to make these zippered iterations more explicit and to permit iteration over a tuple’s values directly. Thus, the zippered iteration above would now be written:
forall (i,a) in zip(1..n, A)
ignoring tuple components/underscore¶
Overtime, the mechanism used to ignore a tuple component when destructuring a tuple
has changed.
Originally, an underscore was used to drop a value on the floor. For example,
given a 3-tuple t
, the first and last components could be stored in x
and z
,
dropping the second component on the floor using: var (x, _, z) = t;
.
In version 1.1 (Apr 2010), we changed this to use a blank space instead of an
underscore, for simplicity and to permit underscore to be used as an identifier
name. Thus, the example above would have been written as var (x, , z) = t;
during this time period.
However, in 2012, we changed back to using the underscore again in order to
support the specification of 1-tuples using a dangling comma, similar to
Python. Thus, dropping a tuple component is expressed as var (x, _, z) = t;
again while (1.2, )
is a 1-tuple of reals.
version 1.4, October 2011¶
function declaration keywords¶
Prior to this change, the keyword def
was used to
define both procedures and iterators; the compiler inspected the body of the
function for yield statements to determine whether it was a procedure or an
iterator.
In this change, we introduced the proc
and iter
keywords to distinguish
between these two cases for the sake of clarity, to avoid mistakes, to support
semantics that clearly distinguish between these cases, and to better support
specifying interfaces.