!standard C.6.3(0) 19-03-11 AI12-0321-1/03
!standard C.6.4(0)
!class Amendment 19-03-07
!status Amendment 1-2012 19-03-11
!status ARG Approved 9-0-1 19-03-11
!status work item 19-03-07
!status received 19-03-07
!priority Low
!difficulty Easy
!subject Support for Arithmetic Atomic Operations and Test and Set
!summary
!problem
On multiprocessor platforms, relying on locks for synchronization can be
problematic. One issue is that while a task is blocked, it can be difficult or
impossible to offer guarantees for progress. Other problems associated with
locks includes deadlock, livelock, and priority inversion. Locking can also be a
significant detriment to performance, and reduce opportunities for parallelism.
Lock-free data structures guarantee system-wide progress, while wait-free data
structures, in addition to being lock-free, also guarantee per-thread progresss.
Lock-free data structures can also be used to improve performance by increasing
the amount of time spent executing in parallel since access to the data
structure does not need to be serialised.
AI12-0234-1 provides access to swap and compare-and-swap operations, but there
are other useful operations that could be provided as atomic primitives.
Atomic arithmetic, such as being able to atomically increment an atomic
counter is a common use for atomic operations. Another common need in this
area is to create spin-locks in user space via test and set instructions.
Ada should provide some simple primitives that can be mapped to hardware
instructions that allow such updates to perform as expected.
!proposal
This proposal depends on AI12-0234-1, which defines the parent package
Ada.Atomic_Operations.
The solution is to provide a set of standard library calls that map to commonly
available atomic hardware instructions such as test and set, and atomic
increment. These subprograms are to be intrinsic calls.
The libraries are generic libraries in order to support operations on discrete
types of different sizes, and we require that the actual types for the generics
be atomic types, so that Ada's semantics of atomic types can be associated
with these primitive operations.
!wording
C.6.3 The Package System.Atomic_Operations.Test_And_Set
The language-defined package System.Atomic_Operations.Test_And_Set
provides an operation to atomically set and clear an atomic flag object.
Static Semantics
The library package System.Atomic_Operations.Test_And_Set has the
following declaration:
package System.Atomic_Operations.Test_And_Set
with Pure, Nonblocking is
type Test_And_Set_Flag is mod
with Atomic, Default_Value => 0, Size => ;
function Atomic_Test_And_Set
(Item : aliased in out Test_And_Set_Flag) return Boolean
with Convention => Intrinsic;
procedure Atomic_Clear
(Item : aliased in out Test_And_Set_Flag)
with Convention => Intrinsic;
function Is_Lock_Free
(Item : aliased Test_And_Set_Flag) return Boolean
with Convention => Intrinsic;
end System.Atomic_Operations.Test_And_Set;
Test_And_Set_Flag represents the state of an atomic flag object.
An atomic flag object can either be considered to be set or cleared.
Atomic_Test_And_Set performs an atomic test-and-set operation on Item.
Item is set to some implementation-defined nonzero value. The function returns
True if the previous contents were nonzero, and otherwise returns False.
Atomic_Clear performs an atomic clear operation on Item. After the
operation, Item contains 0. This call should be used in conjunction with
Atomic_Test_And_Set.
C.6.4 The Package System.Atomic_Operations.Arithmetic
The language-defined generic package System.Atomic_Operations.Arithmetic
provides operations to perform arithmetic atomically on objects of
integer types.
Static Semantics
The generic library package System.Atomic_Operations.Arithmetic has the
following declaration:
generic
type Atomic_Type is range <> with Atomic;
package System.Atomic_Operations.Arithmetic
with Pure, Nonblocking is
procedure Atomic_Add (Item : aliased in out Atomic_Type;
Value : Atomic_Type)
with Convention => Intrinsic;
procedure Atomic_Subtract (Item : aliased in out Atomic_Type;
Value : Atomic_Type)
with Convention => Intrinsic;
function Atomic_Fetch_And_Add
(Item : aliased in out Atomic_Type;
Value : Atomic_Type) return Atomic_Type
with Convention => Intrinsic;
function Atomic_Fetch_And_Subtract
(Item : aliased in out Atomic_Type;
Value : Atomic_Type) return Atomic_Type
with Convention => Intrinsic;
function Is_Lock_Free (Item : aliased Atomic_Type) return Boolean
with Convention => Intrinsic;
end System.Atomic_Operations.Arithmetic;
The operations of this package are defined as follows:
procedure Atomic_Add (Item : aliased in out Atomic_Type;
Value : Atomic_Type)
with Convention => Intrinsic;
Atomically performs: Item := Item + Value;
procedure Atomic_Subtract (Item : aliased in out Atomic_Type;
Value : Atomic_Type)
with Convention => Intrinsic;
Atomically performs: Item := Item - Value;
function Atomic_Fetch_And_Add (Item : aliased in out Atomic_Type;
Value : Atomic_Type) return Atomic_Type
with Convention => Intrinsic;
Atomically performs: Tmp := Item; Item := Item + Value; return Tmp;
function Atomic_Fetch_And_Subtract (Item : aliased in out Atomic_Type;
Value : Atomic_Type) return Atomic_Type
with Convention => Intrinsic;
Atomically performs: Tmp := Item; Item := Item - Value; return Tmp;
!discussion
The approach taken to improving support for lock free data structures is to
provide a set of libraries of low level atomic primitives similar to the library
that is provided by gcc for C and C++.
The library of intrinsic primitives might be of interest to those wishing to
implement specific lock free algorithms, particularly if porting those applications
from other languages.
It was considered whether modular arithmetic functions should be provided.
While it is possible, the routines would be trickier to write because
in the cases where the modulus of the modular types is not a multiple of
System.Storage_Unit. For such cases, updating an atomic variable might
need to include a write back to the the atomic variable to handle the
wrap around. With signed arithmetic, we need to generate Constraint_Error
for overflow checks, so checking for overflow is needed. But if
extra performance is needed, those checks can be suppressed.
!corrigendum C.6.3
@dinsc
The language-defined package System.Atomic_Operations.Test_And_Set
provides an operation to atomically set and clear an atomic flag object.
@s8<@i>
The library package System.Atomic_Operations.Test_And_Set has the
following declaration:
@xcode<@b System.Atomic_Operations.Test_And_Set
@b Pure, Nonblocking @b>
@xcode< @b Test_And_Set_Flag @b @ft<@i>
@b Atomic, Default_Value =@> 0, Size =@> @ft<@i>;>
@xcode< @b Atomic_Test_And_Set
(Item : @b Test_And_Set_Flag) @b Boolean
@b Convention =@> Intrinsic;>
@xcode< @b Atomic_Clear
(Item : @b Test_And_Set_Flag)
@b Convention =@> Intrinsic;>
@xcode< @b Is_Lock_Free
(Item : @b Test_And_Set_Flag) @b Boolean
@b Convention =@> Intrinsic;>
@xcode<@b System.Atomic_Operations.Test_And_Set;>
Test_And_Set_Flag represents the state of an atomic flag object.
An atomic flag object can either be considered to be set or cleared.
Atomic_Test_And_Set performs an atomic test-and-set operation on Item.
Item is set to some implementation-defined nonzero value. The function returns
True if the previous contents were nonzero, and otherwise returns False.
Atomic_Clear performs an atomic clear operation on Item. After the
operation, Item contains 0. This call should be used in conjunction with
Atomic_Test_And_Set.
!corrigendum C.6.4
@dinsc
The language-defined generic package System.Atomic_Operations.Arithmetic
provides operations to perform arithmetic atomically on objects of
integer types.
@s8<@i>
The generic library package System.Atomic_Operations.Arithmetic has the
following declaration:
@xcode<@b
@b Atomic_Type @b <@> @b Atomic;
@b System.Atomic_Operations.Arithmetic
@b Pure, Nonblocking @b>
@xcode< @b Atomic_Add (Item : @b Atomic_Type;
Value : Atomic_Type)
@b Convention =@> Intrinsic;>
@xcode< @b Atomic_Subtract (Item : @b Atomic_Type;
Value : Atomic_Type)
@b Convention =@> Intrinsic;>
@xcode< @b Atomic_Fetch_And_Add
(Item : @b Atomic_Type;
Value : Atomic_Type) @b Atomic_Type
@b Convention =@> Intrinsic;>
@xcode< @b Atomic_Fetch_And_Subtract
(Item : @b Atomic_Type;
Value : Atomic_Type) @b Atomic_Type
@b Convention =@> Intrinsic;>
@xcode< @b Is_Lock_Free (Item : @b Atomic_Type) @b Boolean
@b Convention =@> Intrinsic;>
@xcode<@b System.Atomic_Operations.Arithmetic;>
The operations of this package are defined as follows:
@xcode<@b Atomic_Add (Item : @b Atomic_Type;
Value : Atomic_Type)
@b Convention =@> Intrinsic;>
@xindent>
@xcode<@b Atomic_Subtract (Item : @b Atomic_Type;
Value : Atomic_Type)
@b Convention =@> Intrinsic;>
@xindent>
@xcode<@b Atomic_Fetch_And_Add
(Item : @b Atomic_Type;
Value : Atomic_Type) @b Atomic_Type
@b Convention =@> Intrinsic;>
@xindent Tmp;>>
@xcode<@b Atomic_Fetch_And_Subtract
(Item : @b Atomic_Type;
Value : Atomic_Type) @b Atomic_Type
@b Convention =@> Intrinsic;>
@xindent Tmp;>>
!ASIS
No ASIS effect.
!ACATS test
An ACATS C-Test is needed to check that the new capabilities are supported.
!appendix
From: Brad Moore
Sent: Thursday, March 07, 2019 6:03 PM
Here is my first draft for a new AI on additional atomic operations.
This content was split off from AI12-0234-1.
[This is version /01 of this AI - Editor.]
I eliminated functions that did not seem to be needed, such as atomic load,
atomic store, and atomic fences.
I also eliminated the signed arithmetic package. The modular arithmetic better
maps to the underlying calls, since the signed onces would need to support
detection of arithmetic overflow, which the underlying gcc instrinsic functions
do not do. We could add signed arithmetic functions in the future if the need
was great enough.
****************************************************************
From: Brad Moore
Sent: Saturday, March 09, 2019 1:51 PM
Here is an update to AI12-0321-1 [This is version /02 of the AI - Editor.]
The main changes are;
- Atomic_Exchange was pulled out of this AI, and moved to AI12-234-1
- The child packages of Atomic_Operations, are rooted under package System, rather than package Ada.
- Wording was moved to follow section C.6, Shared Variable Control
- The functions of the form
Atomic_xxx_And_Fetch were removed.
The C standard only defines Atomic_Fetch_And_xxx functions, and it didn't seem
like it was worth providing both forms.
- The following procedures were added to Atomic_Operations.Arithmetic
procedure Atomic_Add (Item : aliased in out Atomic_Type;
Value : Atomic_Type)
with Convention => Intrinsic;
procedure Atomic_Subtract (Item : aliased in out Atomic_Type;
Value : Atomic_Type)
with Convention => Intrinsic;
procedure Atomic_Bitwise_And (Item : aliased in out Atomic_Type;
Value : Atomic_Type)
with Convention => Intrinsic;
procedure Atomic_Bitwise_Or (Item : aliased in out Atomic_Type;
Value : Atomic_Type)
with Convention => Intrinsic;
procedure Atomic_Xor (Item : aliased in out Atomic_Type;
Value : Atomic_Type)
with Convention => Intrinsic;
This is because all the other functions return the original value before the
operation was applied, and in many cases, the user will not need these values.
This can simplify the implementation of the calls, and more optimal performance.
- The call
function Atomic_Fetch_And_Nand (Item : aliased in out Atomic_Type;
Value : Atomic_Type) return Atomic_Type
with Intrinsic;
Was removed. The C11 standard does not define this function, and it doesn't seem
likely that there is a great need to perform atomic nands.
If such a need materializes, we could add it in the future.
- Is_Lock_Free was added to the Test_And_Set child package and the Arithmetic
child package.
- The generic formal type for the Atomic_Operations.Arithmetic package was
changed from a modular generic formal type to a signed arithmetic generic
formal type.
This is because it was perceived that supporting atomic updates for modular
types where the modulus is not a multiple of the storage unit, is trickier, and
possibly requiring loops involving compare and swap.
Using signed arithmetic means there is a need to check for arithmetic overflow,
but that can be done without looping, and if performance is an issue, the
overflow checks can be suppressed.
****************************************************************
From: Brad Moore
Sent: Wednesday, March 13, 2019 12:22 AM
In the recent electronic ARG meeting, we were discussing the
Atomic_Operations.Arithmetic package.
It was mentioned that the Bitwise operations that were defined in AI12-0321-1
did not make sense if the generic formal type is a signed integer type, so we
decided to drop those subprograms from the AI. The thought was that Atomic
Addition and Subtraction is likely the most common need to support,
and that Bitwise operations seemed to be much less of a need.
I mentioned that prior to Saturday, the generic formal of the Arithmetic
package was a modular type, but that I had changed it to signed integer,
due to perceived issues with dealing with addition and subtraction, when the
modulus of the generic formal type was not a multiple of the machines storage
element size.
The decision to have the Arithmetic package only support the Add and Subtract,
still makes sense to me.
But I am now thinking we were a bit too hasty in tossing out the bitwise
subprograms.
After thinking about it, I think atomically modifying and testing bits
atomically is probably something that could be quite useful.
Something I might have used for example, in my Paraffin libraries.
If we wanted to support atomic bitwise operations, I think we'd want to limit
the support to modular types whose modulus is a power of two.
That way, one cannot get into trouble setting or clearing bits of the generic
formal modular type.
All bit patterns (for values less than the 2**modulus) are valid bit patterns,
so there would be no need to add additional checks on the results of these
operations.
Fortunately, that restriction could easily be enforced by placing the
following assert in the generic package.
-- Only Modular types whose modulus is a power of 2 are allowed
pragma Assert ((Atomic_Type'Base (Atomic_Type'Modulus) and
Atomic_Type'Base (Atomic_Type'Modulus - 1)) = 0);
Since adding these functions would be very similar to the wording for adding
the Atomic_Operations.Arithmetic package, I think we should consider adding
the following which supports three operations, "or", "and", and "xor";
generic
type Atomic_Type is mod <> with Atomic;
package System.Atomic_Operations.Bitwise with Pure, Nonblocking
is
-- Only Modular types whose modulus is a power of 2 are allowed
pragma Assert ((Atomic_Type'Base (Atomic_Type'Modulus) and
Atomic_Type'Base (Atomic_Type'Modulus - 1)) = 0);
procedure Atomic_Bitwise_And (Item : aliased in out Atomic_Type;
Value : Atomic_Type)
with Convention => Intrinsic;
procedure Atomic_Bitwise_Or (Item : aliased in out Atomic_Type;
Value : Atomic_Type)
with Convention => Intrinsic;
procedure Atomic_Bitwise_Xor (Item : aliased in out Atomic_Type;
Value : Atomic_Type)
with Convention => Intrinsic;
function Atomic_Fetch_And_Bitwise_And
(Item : aliased in out Atomic_Type;
Value : Atomic_Type) return Atomic_Type
with Convention => Intrinsic;
function Atomic_Fetch_And_Bitwise_Or
(Item : aliased in out Atomic_Type;
Value : Atomic_Type) return Atomic_Type
with Convention => Intrinsic;
function Atomic_Fetch_And_Bitwise_Xor
(Item : aliased in out Atomic_Type;
Value : Atomic_Type) return Atomic_Type
with Convention => Intrinsic;
function Is_Lock_Free
(Item : aliased Atomic_Type) return Boolean
with Convention => Intrinsic, Nonblocking
end System.Atomic_Operations.Bitwise;
I'd be happy to write this up (likely as a new AI if it cant be added to
AI12-0321-1.
Thoughts?
***************************************************************
From: Randy Brukardt
Sent: Wednesday, March 13, 2019 12:49 AM
...
> Fortunately, that restriction could easily be enforced by placing the
> following assert in the generic package.
>
> -- Only Modular types whose modulus is a power of 2 are allowed
> pragma Assert ((Atomic_Type'Base (Atomic_Type'Modulus) and
> Atomic_Type'Base (Atomic_Type'Modulus - 1)) = 0);
I think this Assertion always raises Constraint_Error (or is illegal).
Atomic_Type'Modulus is a universal_integer value, and those do not wrap.
For all T, T'Modulus > T'Last. Ergo, T'Modulus is out of range. If T'Modulus
was static (as it might be in an instance), then the type conversion is
illegal.
You need to somehow use 'Mod. However, T'Mod(T'Modulus) is always 0.
I've usually had to declare a Big_Mod type for this sort of thing:
type Big_Mod is mod System.Max_Binary_Modulus;
and then you can write your expression, so long as you pretest for modulii
that are too large. Which is a royal mess. :-)
Anyway, you need to go back to the drawing board with this assertion.
****************************************************************
From: Brad Moore
Sent: Wednesday, March 13, 2019 1:32 AM
> I think this Assertion always raises Constraint_Error (or is illegal).
> Atomic_Type'Modulus is a universal_integer value, and those do not
> wrap. For all T, T'Modulus > T'Last. Ergo, T'Modulus is out of range.
> If T'Modulus was static (as it might be in an instance), then the type
> conversion is illegal.
I knew that Atomic_Type'Modulus is a universal_integer value, which is why
I converted that result to the Atomic_Type'Base subtype, which allows me to
do the bitwise and operation.
I tried this with GNAT, and it compiles, and works fine in my test program.
When I try it with a type such as
type Mod_5 is mod 5;
It correctly tells me that an exception will be raised at runtime.
(since the modulus of that subtype is not a power of 2)
****************************************************************
From: Tucker Taft
Sent: Wednesday, March 13, 2019 8:40 AM
In any case, I think we should leave these out for now, as they are less
critical than add/subtract.
It is easy enough to write such a package yourself if you know what you are
doing. Here we are looking for widely useful capabilities that need
standardization.
****************************************************************
From: Randy Brukardt
Sent: Wednesday, March 13, 2019 3:41 PM
> I knew that Atomic_Type'Modulus is a universal_integer value, which is
> why I converted that result to the Atomic_Type'Base subtype, which
> allows me to do the bitwise and operation.
Yes, and that conversion always raises Constraint_Error (or is illegal if the
type is static).
> I tried this with GNAT, and it compiles, and works fine in my test
> program.
Either you didn't actually test what you think you did, or GNAT is wrong.
4.6(28) clearly states that Constraint_Error is raised for a value outside of
the base range of a modular type. T'Modulus is by definition outside of the
base range of T. And 4.9 states that a static expression is illegal if it
raises any exception other than an overflow.
Now, I know that if you compiled this with Janus/Ada, it would in fact work
with at least some types, but that would be because Janus/Ada doesn't
properly handle these conversions when found in a generic. In particular,
T'Modulus of mod 2**32 has the special value zero at runtime so we can handle
such types correctly, but that would of course cause oddities if actually used
explicitly. But just because it would work doesn't make it right.
So I wouldn't be surprised if it accidentally worked for Max_Binary_Modulus,
but I wouldn't take that as some sort of assumption that it should work. I'd
suggest trying it in a non-generic case to see.
I'm pretty sure that there is an ACATS test that checks this rule, but perhaps
it could use some beefing up.
> When I try it with a type such as
>
> type Mod_5 is mod 5;
>
> It correctly tells me that an exception will be raised at runtime.
> (since the modulus of that subtype is not a power of 2)
Surely this is true, since it *always* raises Constraint_Error.
****************************************************************
From: Tucker Taft
Sent: Wednesday, March 13, 2019 4:30 PM
Another test that should work, I believe, is:
pragma Assert (T'Modulus = 2 ** T'Size);
unless the 'Size has been explicitly specified to be larger.
Or if you want to go whole hog:
pragma Assert((for some I in 1..T'Size => T'Modulus = 2**I));
though now you are getting into non-static expressions, so Constraint_Error
becomes more likely if you bump up against the Max_Integer limits of run-time
universal-int computations.
;-)
****************************************************************
From: Bob Duff
Sent: Wednesday, March 13, 2019 6:00 PM
> pragma Assert((for some I in 1..T'Size => T'Modulus = 2**I));
For "type T is mod 1;", T'Size = 0, 2**T'Size = 1, and T'Modulus = 1.
;-)
Note to Brad: Unlike signed integers, the base range of a modular type is
widened beyond the first subtype.
****************************************************************
From: Tucker Taft
Sent: Wednesday, March 13, 2019 7:27 PM
Yes, I realized that, but I was somehow unable to allow myself to include the
"mod 1" option. ;-)
****************************************************************
From: Tucker Taft
Sent: Wednesday, March 13, 2019 7:33 PM
> Note to Brad: Unlike signed integers, the base range of a modular
> type is widened beyond the first subtype.
What does this mean? Did you mean the base range is *not* widened?
****************************************************************
From: Bob Duff
Sent: Wednesday, March 13, 2019 8:12 PM
Yes, thanks for the correction.
****************************************************************
From: Brad Moore
Sent: Wednesday, March 13, 2019 11:30 PM
Thanks for the explanations.
By the way, the reason I thought the following example was OK, is because
I didn't have the "enable assertions" checkbox checked, for the project file.
Oddly, I do get a warning if I uncomment the assert below, which says that
an assertion would fail at run time. But I don't get any warnings otherwise,
and the code runs successfully.
If I enable assertions, then the code does not compile, as Randy said.
with Ada.Text_IO;
use Ada.Text_IO;
procedure Main is
type Mod_8 is mod 2**3;
type Mod_5 is mod 5;
type Mod_16 is mod 2**4;
type Mod_32 is mod 2**5;
pragma Assert ((Mod_8'Base (Mod_8'Modulus) and
Mod_8'Base (Mod_8'Modulus - 1)) = 0);
pragma Assert ((Mod_16'Base (Mod_16'Modulus) and
Mod_16'Base (Mod_16'Modulus - 1)) = 0);
pragma Assert ((Mod_32'Base (Mod_32'Modulus) and
Mod_32'Base (Mod_32'Modulus - 1)) = 0);
-- pragma Assert (((Mod_5'Base (Mod_5'Modulus) and
-- Mod_5'Base (Mod_5'Modulus - 1)) = 0));
X : Mod_8;
X2 : Mod_16;
X3 : Mod_32;
begin
X := 4;
X := X + 7;
X2 := 4;
X2 := X2 + 14;
X3 := 5;
X3 := X3 + 30;
Put_Line ("Mod8:" & Mod_8'Image (X));
Put_Line ("Mod16:" & Mod_16'Image (X2));
Put_Line ("Mod32:" & Mod_32'Image (X3));
end Main;
****************************************************************
From: Brad Moore
Sent: Wednesday, March 13, 2019 11:39 PM
Thanks for the suggestion, that works (the first simpler version, at least)
for my example. A nice concise way to express this property.
****************************************************************
From: Randy Brukardt
Sent: Thursday, March 14, 2019 7:21 PM
> > pragma Assert((for some I in 1..T'Size => T'Modulus = 2**I));
>
> For "type T is mod 1;", T'Size = 0, 2**T'Size = 1, and T'Modulus = 1.
>
> ;-)
Tucker noted that it isn't a very good assertion, since someone could
declare:
type Biggie is mod 256 with Atomic, Size => 16;
OTOH, we could decide to explicitly not care about such corner cases.
> Note to Brad: Unlike signed integers, the base range of a modular
> type is widened beyond the first subtype.
Every time I read this, I'm confused. I even went and looked it up in the RM
to be sure. I think there is a "not" missing somewhere. Perhaps you meant:
Note to Brad: Unlike signed integers, the base range of a
modular type is NOT widened beyond the first subtype.
****************************************************************