15. Implementation of Specific Ada Features

This chapter describes the GNAT implementation of several Ada language facilities.

15.1. Machine Code Insertions

Package Machine_Code provides machine code support as described in the Ada Reference Manual in two separate forms:

  • Machine code statements, consisting of qualified expressions that fit the requirements of RM section 13.8.
  • An intrinsic callable procedure, providing an alternative mechanism of including machine instructions in a subprogram.

The two features are similar, and both are closely related to the mechanism provided by the asm instruction in the GNU C compiler. Full understanding and use of the facilities in this package requires understanding the asm instruction, see the section on Extended Asm in Using_the_GNU_Compiler_Collection_(GCC).

Calls to the function Asm and the procedure Asm have identical semantic restrictions and effects as described below. Both are provided so that the procedure call can be used as a statement, and the function call can be used to form a code_statement.

Consider this C asm instruction:

asm ("fsinx %1 %0" : "=f" (result) : "f" (angle));

The equivalent can be written for GNAT as:

Asm ("fsinx %1 %0",
     My_Float'Asm_Output ("=f", result),
     My_Float'Asm_Input  ("f",  angle));

The first argument to Asm is the assembler template, and is identical to what is used in GNU C. This string must be a static expression. The second argument is the output operand list. It is either a single Asm_Output attribute reference, or a list of such references enclosed in parentheses (technically an array aggregate of such references).

The Asm_Output attribute denotes a function that takes two parameters. The first is a string, the second is the name of a variable of the type designated by the attribute prefix. The first (string) argument is required to be a static expression and designates the constraint (see the section on Constraints in Using_the_GNU_Compiler_Collection_(GCC)) for the parameter; e.g., what kind of register is required. The second argument is the variable to be written or updated with the result. The possible values for constraint are the same as those used in the RTL, and are dependent on the configuration file used to build the GCC back end. If there are no output operands, then this argument may either be omitted, or explicitly given as No_Output_Operands. No support is provided for GNU C’s symbolic names for output parameters.

The second argument of my_float'Asm_Output functions as though it were an out parameter, which is a little curious, but all names have the form of expressions, so there is no syntactic irregularity, even though normally functions would not be permitted out parameters. The third argument is the list of input operands. It is either a single Asm_Input attribute reference, or a list of such references enclosed in parentheses (technically an array aggregate of such references).

The Asm_Input attribute denotes a function that takes two parameters. The first is a string, the second is an expression of the type designated by the prefix. The first (string) argument is required to be a static expression, and is the constraint for the parameter, (e.g., what kind of register is required). The second argument is the value to be used as the input argument. The possible values for the constraint are the same as those used in the RTL, and are dependent on the configuration file used to built the GCC back end. No support is provided for GNU C’s symbolic names for input parameters.

If there are no input operands, this argument may either be omitted, or explicitly given as No_Input_Operands. The fourth argument, not present in the above example, is a list of register names, called the clobber argument. This argument, if given, must be a static string expression, and is a space or comma separated list of names of registers that must be considered destroyed as a result of the Asm call. If this argument is the null string (the default value), then the code generator assumes that no additional registers are destroyed. In addition to registers, the special clobbers memory and cc as described in the GNU C docs are both supported.

The fifth argument, not present in the above example, called the volatile argument, is by default False. It can be set to the literal value True to indicate to the code generator that all optimizations with respect to the instruction specified should be suppressed, and in particular an instruction that has outputs will still be generated, even if none of the outputs are used. See Using_the_GNU_Compiler_Collection_(GCC) for the full description. Generally it is strongly advisable to use Volatile for any ASM statement that is missing either input or output operands or to avoid unwanted optimizations. A warning is generated if this advice is not followed.

No support is provided for GNU C’s asm goto feature.

The Asm subprograms may be used in two ways. First the procedure forms can be used anywhere a procedure call would be valid, and correspond to what the RM calls ‘intrinsic’ routines. Such calls can be used to intersperse machine instructions with other Ada statements. Second, the function forms, which return a dummy value of the limited private type Asm_Insn, can be used in code statements, and indeed this is the only context where such calls are allowed. Code statements appear as aggregates of the form:

Asm_Insn'(Asm (...));
Asm_Insn'(Asm_Volatile (...));

In accordance with RM rules, such code statements are allowed only within subprograms whose entire body consists of such statements. It is not permissible to intermix such statements with other Ada statements.

Typically the form using intrinsic procedure calls is more convenient and more flexible. The code statement form is provided to meet the RM suggestion that such a facility should be made available. The following is the exact syntax of the call to Asm. As usual, if named notation is used, the arguments may be given in arbitrary order, following the normal rules for use of positional and named arguments:

ASM_CALL ::= Asm (
                 [Template =>] static_string_EXPRESSION
               [,[Outputs  =>] OUTPUT_OPERAND_LIST      ]
               [,[Inputs   =>] INPUT_OPERAND_LIST       ]
               [,[Clobber  =>] static_string_EXPRESSION ]
               [,[Volatile =>] static_boolean_EXPRESSION] )


  SUBTYPE_MARK'Asm_Output (static_string_EXPRESSION, NAME)



The identifiers No_Input_Operands and No_Output_Operands are declared in the package Machine_Code and must be referenced according to normal visibility rules. In particular if there is no use clause for this package, then appropriate package name qualification is required.

15.2. GNAT Implementation of Tasking

This chapter outlines the basic GNAT approach to tasking (in particular, a multi-layered library for portability) and discusses issues related to compliance with the Real-Time Systems Annex.

15.2.1. Mapping Ada Tasks onto the Underlying Kernel Threads

GNAT’s run-time support comprises two layers:

  • GNARL (GNAT Run-time Layer)
  • GNULL (GNAT Low-level Library)

In GNAT, Ada’s tasking services rely on a platform and OS independent layer known as GNARL. This code is responsible for implementing the correct semantics of Ada’s task creation, rendezvous, protected operations etc.

GNARL decomposes Ada’s tasking semantics into simpler lower level operations such as create a thread, set the priority of a thread, yield, create a lock, lock/unlock, etc. The spec for these low-level operations constitutes GNULLI, the GNULL Interface. This interface is directly inspired from the POSIX real-time API.

If the underlying executive or OS implements the POSIX standard faithfully, the GNULL Interface maps as is to the services offered by the underlying kernel. Otherwise, some target dependent glue code maps the services offered by the underlying kernel to the semantics expected by GNARL.

Whatever the underlying OS (VxWorks, UNIX, Windows, etc.) the key point is that each Ada task is mapped on a thread in the underlying kernel. For example, in the case of VxWorks, one Ada task = one VxWorks task.

In addition Ada task priorities map onto the underlying thread priorities. Mapping Ada tasks onto the underlying kernel threads has several advantages:

  • The underlying scheduler is used to schedule the Ada tasks. This makes Ada tasks as efficient as kernel threads from a scheduling standpoint.
  • Interaction with code written in C containing threads is eased since at the lowest level Ada tasks and C threads map onto the same underlying kernel concept.
  • When an Ada task is blocked during I/O the remaining Ada tasks are able to proceed.
  • On multiprocessor systems Ada tasks can execute in parallel.

Some threads libraries offer a mechanism to fork a new process, with the child process duplicating the threads from the parent. GNAT does not support this functionality when the parent contains more than one task.

15.2.2. Ensuring Compliance with the Real-Time Annex

Although mapping Ada tasks onto the underlying threads has significant advantages, it does create some complications when it comes to respecting the scheduling semantics specified in the real-time annex (Annex D).

For instance the Annex D requirement for the FIFO_Within_Priorities scheduling policy states:

When the active priority of a ready task that is not running changes, or the setting of its base priority takes effect, the task is removed from the ready queue for its old active priority and is added at the tail of the ready queue for its new active priority, except in the case where the active priority is lowered due to the loss of inherited priority, in which case the task is added at the head of the ready queue for its new active priority.

While most kernels do put tasks at the end of the priority queue when a task changes its priority, (which respects the main FIFO_Within_Priorities requirement), almost none keep a thread at the beginning of its priority queue when its priority drops from the loss of inherited priority.

As a result most vendors have provided incomplete Annex D implementations.

The GNAT run-time, has a nice cooperative solution to this problem which ensures that accurate FIFO_Within_Priorities semantics are respected.

The principle is as follows. When an Ada task T is about to start running, it checks whether some other Ada task R with the same priority as T has been suspended due to the loss of priority inheritance. If this is the case, T yields and is placed at the end of its priority queue. When R arrives at the front of the queue it executes.

Note that this simple scheme preserves the relative order of the tasks that were ready to execute in the priority queue where R has been placed at the end.

15.2.3. Support for Locking Policies

This section specifies which policies specified by pragma Locking_Policy are supported on which platforms.

GNAT supports the standard Ceiling_Locking policy, and the implementation defined Inheritance_Locking and Concurrent_Readers_Locking policies.

Ceiling_Locking is supported on all platforms if the operating system supports it. In particular, Ceiling_Locking is not supported on VxWorks. Inheritance_Locking is supported on Linux, Darwin (Mac OS X), LynxOS 178, and VxWorks. Concurrent_Readers_Locking is supported on Linux.

Notes about Ceiling_Locking on Linux: If the process is running as ‘root’, ceiling locking is used. If the capabilities facility is installed (“sudo apt-get –assume-yes install libcap-dev” on Ubuntu, for example), and the program is linked against that library (“-largs -lcap”), and the executable file has the cap_sys_nice capability (“sudo /sbin/setcap cap_sys_nice=ep executable_file_name”), then ceiling locking is used. Otherwise, the Ceiling_Locking policy is ignored.

15.3. GNAT Implementation of Shared Passive Packages

GNAT fully implements the pragma Shared_Passive for the purpose of designating shared passive packages. This allows the use of passive partitions in the context described in the Ada Reference Manual; i.e., for communication between separate partitions of a distributed application using the features in Annex E.

However, the implementation approach used by GNAT provides for more extensive usage as follows:

Communication between separate programs
This allows separate programs to access the data in passive partitions, using protected objects for synchronization where needed. The only requirement is that the two programs have a common shared file system. It is even possible for programs running on different machines with different architectures (e.g., different endianness) to communicate via the data in a passive partition.
Persistence between program runs
The data in a passive package can persist from one run of a program to another, so that a later program sees the final values stored by a previous run of the same program.

The implementation approach used is to store the data in files. A separate stream file is created for each object in the package, and an access to an object causes the corresponding file to be read or written.

The environment variable SHARED_MEMORY_DIRECTORY should be set to the directory to be used for these files. The files in this directory have names that correspond to their fully qualified names. For example, if we have the package

package X is
  pragma Shared_Passive (X);
  Y : Integer;
  Z : Float;
end X;

and the environment variable is set to /stemp/, then the files created will have the names:


These files are created when a value is initially written to the object, and the files are retained until manually deleted. This provides the persistence semantics. If no file exists, it means that no partition has assigned a value to the variable; in this case the initial value declared in the package will be used. This model ensures that there are no issues in synchronizing the elaboration process, since elaboration of passive packages elaborates the initial values, but does not create the files.

The files are written using normal Stream_IO access. If you want to be able to communicate between programs or partitions running on different architectures, then you should use the XDR versions of the stream attribute routines, since these are architecture independent.

If active synchronization is required for access to the variables in the shared passive package, then as described in the Ada Reference Manual, the package may contain protected objects used for this purpose. In this case a lock file (whose name is ___lock (three underscores) is created in the shared memory directory.

This is used to provide the required locking semantics for proper protected object synchronization.

GNAT supports shared passive packages on all platforms except for OpenVMS.

15.4. Code Generation for Array Aggregates

Aggregates have a rich syntax and allow the user to specify the values of complex data structures by means of a single construct. As a result, the code generated for aggregates can be quite complex and involve loops, case statements and multiple assignments. In the simplest cases, however, the compiler will recognize aggregates whose components and constraints are fully static, and in those cases the compiler will generate little or no executable code. The following is an outline of the code that GNAT generates for various aggregate constructs. For further details, you will find it useful to examine the output produced by the -gnatG flag to see the expanded source that is input to the code generator. You may also want to examine the assembly code generated at various levels of optimization.

The code generated for aggregates depends on the context, the component values, and the type. In the context of an object declaration the code generated is generally simpler than in the case of an assignment. As a general rule, static component values and static subtypes also lead to simpler code.

15.4.1. Static constant aggregates with static bounds

For the declarations:

type One_Dim is array (1..10) of integer;
ar0 : constant One_Dim := (1, 2, 3, 4, 5, 6, 7, 8, 9, 0);

GNAT generates no executable code: the constant ar0 is placed in static memory. The same is true for constant aggregates with named associations:

Cr1 : constant One_Dim := (4 => 16, 2 => 4, 3 => 9, 1 => 1, 5 .. 10 => 0);
Cr3 : constant One_Dim := (others => 7777);

The same is true for multidimensional constant arrays such as:

type two_dim is array (1..3, 1..3) of integer;
Unit : constant two_dim := ( (1,0,0), (0,1,0), (0,0,1));

The same is true for arrays of one-dimensional arrays: the following are static:

type ar1b  is array (1..3) of boolean;
type ar_ar is array (1..3) of ar1b;
None  : constant ar1b := (others => false);     --  fully static
None2 : constant ar_ar := (1..3 => None);       --  fully static

However, for multidimensional aggregates with named associations, GNAT will generate assignments and loops, even if all associations are static. The following two declarations generate a loop for the first dimension, and individual component assignments for the second dimension:

Zero1: constant two_dim := (1..3 => (1..3 => 0));
Zero2: constant two_dim := (others => (others => 0));

15.4.2. Constant aggregates with unconstrained nominal types

In such cases the aggregate itself establishes the subtype, so that associations with others cannot be used. GNAT determines the bounds for the actual subtype of the aggregate, and allocates the aggregate statically as well. No code is generated for the following:

type One_Unc is array (natural range <>) of integer;
Cr_Unc : constant One_Unc := (12,24,36);

15.4.3. Aggregates with static bounds

In all previous examples the aggregate was the initial (and immutable) value of a constant. If the aggregate initializes a variable, then code is generated for it as a combination of individual assignments and loops over the target object. The declarations

Cr_Var1 : One_Dim := (2, 5, 7, 11, 0, 0, 0, 0, 0, 0);
Cr_Var2 : One_Dim := (others > -1);

generate the equivalent of

Cr_Var1 (1) := 2;
Cr_Var1 (2) := 3;
Cr_Var1 (3) := 5;
Cr_Var1 (4) := 11;

for I in Cr_Var2'range loop
   Cr_Var2 (I) := -1;
end loop;

15.4.4. Aggregates with nonstatic bounds

If the bounds of the aggregate are not statically compatible with the bounds of the nominal subtype of the target, then constraint checks have to be generated on the bounds. For a multidimensional array, constraint checks may have to be applied to sub-arrays individually, if they do not have statically compatible subtypes.

15.4.5. Aggregates in assignment statements

In general, aggregate assignment requires the construction of a temporary, and a copy from the temporary to the target of the assignment. This is because it is not always possible to convert the assignment into a series of individual component assignments. For example, consider the simple case:

A := (A(2), A(1));

This cannot be converted into:

A(1) := A(2);
A(2) := A(1);

So the aggregate has to be built first in a separate location, and then copied into the target. GNAT recognizes simple cases where this intermediate step is not required, and the assignments can be performed in place, directly into the target. The following sufficient criteria are applied:

  • The bounds of the aggregate are static, and the associations are static.
  • The components of the aggregate are static constants, names of simple variables that are not renamings, or expressions not involving indexed components whose operands obey these rules.

If any of these conditions are violated, the aggregate will be built in a temporary (created either by the front-end or the code generator) and then that temporary will be copied onto the target.

15.5. The Size of Discriminated Records with Default Discriminants

If a discriminated type T has discriminants with default values, it is possible to declare an object of this type without providing an explicit constraint:

type Size is range 1..100;

type Rec (D : Size := 15) is record
   Name : String (1..D);
end T;

Word : Rec;

Such an object is said to be unconstrained. The discriminant of the object can be modified by a full assignment to the object, as long as it preserves the relation between the value of the discriminant, and the value of the components that depend on it:

Word := (3, "yes");

Word := (5, "maybe");

Word := (5, "no"); -- raises Constraint_Error

In order to support this behavior efficiently, an unconstrained object is given the maximum size that any value of the type requires. In the case above, Word has storage for the discriminant and for a String of length 100. It is important to note that unconstrained objects do not require dynamic allocation. It would be an improper implementation to place on the heap those components whose size depends on discriminants. (This improper implementation was used by some Ada83 compilers, where the Name component above would have been stored as a pointer to a dynamic string). Following the principle that dynamic storage management should never be introduced implicitly, an Ada compiler should reserve the full size for an unconstrained declared object, and place it on the stack.

This maximum size approach has been a source of surprise to some users, who expect the default values of the discriminants to determine the size reserved for an unconstrained object: “If the default is 15, why should the object occupy a larger size?” The answer, of course, is that the discriminant may be later modified, and its full range of values must be taken into account. This is why the declaration:

type Rec (D : Positive := 15) is record
   Name : String (1..D);
end record;

Too_Large : Rec;

is flagged by the compiler with a warning: an attempt to create Too_Large will raise Storage_Error, because the required size includes Positive'Last bytes. As the first example indicates, the proper approach is to declare an index type of ‘reasonable’ range so that unconstrained objects are not too large.

One final wrinkle: if the object is declared to be aliased, or if it is created in the heap by means of an allocator, then it is not unconstrained: it is constrained by the default values of the discriminants, and those values cannot be modified by full assignment. This is because in the presence of aliasing all views of the object (which may be manipulated by different tasks, say) must be consistent, so it is imperative that the object, once created, remain invariant.

15.6. Strict Conformance to the Ada Reference Manual

The dynamic semantics defined by the Ada Reference Manual impose a set of run-time checks to be generated. By default, the GNAT compiler will insert many run-time checks into the compiled code, including most of those required by the Ada Reference Manual. However, there are two checks that are not enabled in the default mode for efficiency reasons: checks for access before elaboration on subprogram calls, and stack overflow checking (most operating systems do not perform this check by default).

Strict conformance to the Ada Reference Manual can be achieved by adding two compiler options for dynamic checks for access-before-elaboration on subprogram calls and generic instantiations (-gnatE), and stack overflow checking (-fstack-check).

Note that the result of a floating point arithmetic operation in overflow and invalid situations, when the Machine_Overflows attribute of the result type is False, is to generate IEEE NaN and infinite values. This is the case for machines compliant with the IEEE floating-point standard, but on machines that are not fully compliant with this standard, such as Alpha, the -mieee compiler flag must be used for achieving IEEE confirming behavior (although at the cost of a significant performance penalty), so infinite and NaN values are properly generated.