llvm-capstone/flang/docs/OpenMP-semantics.md
2023-05-10 19:17:47 -05:00

21 KiB
Raw Blame History

OpenMP Semantic Analysis

.. contents::
   :local:

OpenMP for F18

  1. Define and document the parse tree representation for
    • Directives (listed below)
    • Clauses (listed below)
    • Documentation
  2. All the directives and clauses need source provenance for messages
  3. Define and document how an OpenMP directive in the parse tree will be represented as the parent of the statement(s) to which the directive applies. The parser itself will not be able to construct this representation; there will be subsequent passes that do so just like for example do-stmt and do-construct.
  4. Define and document the symbol table extensions
  5. Define and document the module file extensions

Directives

OpenMP divides directives into three categories as follows. The directives that are in the same categories share some characteristics.

Declarative directives

An OpenMP directive may only be placed in a declarative context. A declarative directive results in one or more declarations only; it is not associated with the immediate execution of any user code.

List of existing ones:

  • declare simd
  • declare target
  • threadprivate
  • declare reduction
  • requires

There is a parser node for each of these directives and the parser node saves information associated with the directive, for example, the name of the procedure-name in the declare simd directive.

Each parse tree node keeps source provenance, one for the directive name itself and one for the entire directive starting from the directive name.

A top-level class, OpenMPDeclarativeConstruct, holds all four of the node types as discriminated unions along with the source provenance for the entire directive starting from !$OMP.

In parser-tree.h, OpenMPDeclarativeConstruct is part of the SpecificationConstruct and SpecificationPart in F18 because a declarative directive can only be placed in the specification part of a Fortran program.

All the Names or Designators associated with the declarative directive will be resolved in later phases.

Executable directives

An OpenMP directive that is not declarative. That is, it may only be placed in an executable context. It contains stand-alone directives and constructs that are associated with code blocks. The stand-alone directive is described in the next section.

The constructs associated with code blocks listed below share a similar structure: Begin Directive, Clause List, Code Block, End Directive. The End Directive is optional for constructs like Loop-associated constructs.

  • Block-associated constructs (OpenMPBlockConstruct)
  • Loop-associated constructs (OpenMPLoopConstruct)
  • Atomic construct (OpenMPAtomicConstruct)
  • Sections Construct (OpenMPSectionsConstruct, contains Sections/Parallel Sections constructs)
  • Critical Construct (OpenMPCriticalConstruct)

A top-level class, OpenMPConstruct, includes stand-alone directive and constructs listed above as discriminated unions.

In the parse-tree.h, OpenMPConstruct is an element of the ExecutableConstruct.

All the Names or Designators associated with the executable directive will be resolved in Semantic Analysis.

When the backtracking parser can not identify the associated code blocks, the parse tree will be rewritten later in the Semantics Analysis.

Stand-alone Directives

An OpenMP executable directive that has no associated user code except for that which appears in clauses in the directive.

List of existing ones:

  • taskyield
  • barrier
  • taskwait
  • target enter data
  • target exit data
  • target update
  • ordered
  • flush
  • cancel
  • cancellation point

A higher-level class is created for each category which contains directives listed above that share a similar structure:

  • OpenMPSimpleStandaloneConstruct (taskyield, barrier, taskwait, target enter/exit data, target update, ordered)
  • OpenMPFlushConstruct
  • OpenMPCancelConstruct
  • OpenMPCancellationPointConstruct

A top-level class, OpenMPStandaloneConstruct, holds all four of the node types as discriminated unions along with the source provenance for the entire directive. Also, each parser node for the stand-alone directive saves the source provenance for the directive name itself.

Clauses

Each clause represented as a distinct class in parse-tree.h. A top-level class, OmpClause, includes all the clauses as discriminated unions. The parser node for OmpClause saves the source provenance for the entire clause.

All the Names or Designators associated with the clauses will be resolved in Semantic Analysis.

Note that the backtracking parser will not validate that the list of clauses associated with a directive is valid other than to make sure they are well-formed. In particular, the parser does not check that the association between directive and clauses is correct nor check that the values in the directives or clauses are correct. These checks are deferred to later phases of semantics to simplify the parser.

Symbol Table Extensions for OpenMP

Name resolution can be impacted by the OpenMP code. In addition to the regular steps to do the name resolution, new scopes and symbols may need to be created when encountering certain OpenMP constructs. This section describes the extensions for OpenMP during Symbol Table construction.

OpenMP uses the fork-join model of parallel execution and all OpenMP threads have access to a shared memory place to store and retrieve variables but each thread can also have access to its threadprivate memory that must not be accessed by other threads.

For the directives and clauses that can control the data environments, compiler needs to determine two kinds of access to variables used in the directives associated structured block: shared and private. Each variable referenced in the structured block has an original variable immediately outside of the OpenMP constructs. Reference to a shared variable in the structured block becomes a reference to the original variable. However, each private variable referenced in the structured block, a new version of the original variable (of the same type and size) will be created in the threadprivate memory.

There are exceptions that directives/clauses need to create a new Symbol without creating a new Scope, but in general, when encountering each of the data environment controlling directives (discussed in the following sections), a new Scope will be created. For each private variable referenced in the structured block, a new Symbol is created out of the original variable and the new Symbol is associated with original variables Symbol via HostAssocDetails. A new set of OpenMP specific flags are added into Flag class in symbol.h to indicate the types of associations, data-sharing attributes, and data-mapping attributes in the OpenMP data environments.

New Symbol without new Scope

OpenMP directives that require new Symbol to be created but not new Scope are listed in the following table in terms of the Symbol Table extensions for OpenMP:

Directives/Clauses Create New

Symbol

w/

Add Flag
on Symbol of Flag
Declarative Directives declare simd [(proc-name)] - The name of the enclosing function, subroutine, or interface body to which it applies, or proc-name OmpDeclareSimd
declare target - The name of the enclosing function, subroutine, or interface body to which it applies OmpDeclareTarget
threadprivate(list) - named variables and named common blocks OmpThreadPrivate
declare reduction * reduction-identifier OmpDeclareReduction
Stand-alone directives flush - variable, array section or common block name OmpFlushed
critical [(name)] - name (user-defined identifier) OmpCriticalLock
if ([ directive-name-modifier :] scalar-logical-expr) - directive-name-modifier OmpIfSpecified
  -      No Action

  *      Discussed in “Module File Extensions for OpenMP” section

New Symbol with new Scope

For the following OpenMP regions:

  • target regions
  • target data regions
  • teams regions
  • parallel regions
  • simd regions
  • task generating regions (created by task or taskloop constructs)
  • worksharing regions (created by do, sections, single, or workshare constructs)

A new Scope will be created when encountering the above OpenMP constructs to ensure the correct data environment during the Code Generation. To determine whether a variable referenced in these regions needs the creation of a new Symbol, all the data-sharing attribute rules described in OpenMP Spec [2.15.1] apply during the Name Resolution. The available data-sharing attributes are: shared, private, linear, firstprivate, and lastprivate. The attribute is represented as Flag in the Symbol object.

More details are listed in the following table:

Attribute Create New Symbol Add Flag
on Symbol of Flag
shared No Original variable OmpShared
private Yes New Symbol OmpPrivate
linear Yes New Symbol OmpLinear
firstprivate Yes New Symbol OmpFirstPrivate
lastprivate Yes New Symbol OmpLastPrivate
use_device_ptr Yes New Symbol OmpUseDevicePtr
use_device_addr Yes New Symbol OmpUseDeviceAddr

To determine the right data-sharing attribute, OpenMP defines that the data-sharing attributes of variables that are referenced in a construct can be predetermined, explicitly determined, or implicitly determined.

Predetermined data-sharing attributes

  • Assumed-size arrays are shared
  • The loop iteration variable(s) in the associated do-loop(s) of a do, parallel do, taskloop, or distributeconstruct is (are) private
  • A loop iteration variable for a sequential loop in a parallel or task generating construct is private in the innermost such construct that encloses the loop
  • Implied-do indices and forall indices are private
  • The loop iteration variable in the associated do-loop of a simd construct with just one associated do-loop is linear with a linear-step that is the increment of the associated do-loop
  • The loop iteration variables in the associated do-loop(s) of a simd construct with multiple associated do-loop(s) are lastprivate

Explicitly determined data-sharing attributes

Variables with explicitly determined data-sharing attributes are:

  • Variables are referenced in a given construct
  • Variables are listed in a data-sharing attribute clause on the construct.

The data-sharing attribute clauses are:

  • default clause (discussed in “Implicitly determined data-sharing attributes”)
  • shared clause
  • private clause
  • linear clause
  • firstprivate clause
  • lastprivate clause
  • reduction clause (new Symbol created with the flag OmpReduction set)

Note that variables with predetermined data-sharing attributes may not be listed (with exceptions) in data-sharing attribute clauses.

Implicitly determined data-sharing attributes

Variables with implicitly determined data-sharing attributes are:

  • Variables are referenced in a given construct
  • Variables do not have predetermined data-sharing attributes
  • Variables are not listed in a data-sharing attribute clause on the construct.

Rules for variables with implicitly determined data-sharing attributes:

  • In a parallel construct, if no default clause is present, these variables are shared
  • In a task generating construct, if no default clause is present, a variable for which the data-sharing attribute is not determined by the rules above and that in the enclosing context is determined to be shared by all implicit tasks bound to the current team is shared
  • In a target construct, variables that are not mapped after applying data-mapping attribute rules (discussed later) are firstprivate
  • In an orphaned task generating construct, if no default clause is present, dummy arguments are firstprivate
  • In a task generating construct, if no default clause is present, a variable for which the data-sharing attribute is not determined by the rules above is firstprivate
  • For constructs other than task generating constructs or target constructs, if no default clause is present, these variables reference the variables with the same names that exist in the enclosing context
  • In a parallel, teams, or task generating construct, the data-sharing attributes of these variables are determined by the default clause, if present:
    • default(shared) clause causes all variables referenced in the construct that have implicitly determined data-sharing attributes to be shared
    • default(private) clause causes all variables referenced in the construct that have implicitly determined data-sharing attributes to be private
    • default(firstprivate) clause causes all variables referenced in the construct that have implicitly determined data-sharing attributes to be firstprivate
    • default(none) clause requires that each variable that is referenced in the construct, and that does not have a predetermined data-sharing attribute, must have its data-sharing attribute explicitly determined by being listed in a data-sharing attribute clause

Data-mapping Attribute

When encountering the target data and target directives, the data-mapping attributes of any variable referenced in a target region will be determined and represented as Flag in the Symbol object of the variable. No Symbol or Scope will be created.

However, there are some exceptions for this, Pointers that appear in a use_device_ptr clause are privatized and the device pointers to the corresponding list items in the device data environment are assigned into the private versions so it is best to follow the representation for privatised variables i.e represent them with a new Symbol and OmpUseDevicePtr flag. If a list item that appears in a use_device_addr clause has corresponding storage in the device data environment, references to the list item in the associated structured block are converted into references to the corresponding list item so following the same i.e. represent them with a new Symbol and OmpUseDeviceAddr flag.

The basic steps to determine the data-mapping attribute are:

  1. If map clause is present, the data-mapping attribute is determined by the map-type on the clause and its corresponding Flag are listed below:
data-mapping attribute Flag
to OmpMapTo
from OmpMapFrom
tofrom (default if map-type is not present) OmpMapTo & OmpMapFrom
alloc OmpMapAlloc
release OmpMapRelease
delete OmpMapDelete
  1. Otherwise, the following data-mapping rules apply for variables referenced in a target construct that are not declared in the construct and do not appear in data-sharing attribute or map clauses:
    • If a variable appears in a to or link clause on a declare target directive then it is treated as if it had appeared in a map clause with a map-type of tofrom
  2. Otherwise, the following implicit data-mapping attribute rules apply:
    • If a defaultmap(tofrom:scalar) clause is not present then a scalar variable is not mapped, but instead has an implicit data-sharing attribute of firstprivate
    • If a defaultmap(tofrom:scalar) clause is present then a scalar variable is treated as if it had appeared in a map clause with a map-type of tofrom
    • If a variable is not a scalar then it is treated as if it had appeared in a map clause with a map-type of tofrom

After the completion of the Name Resolution phase, all the data-sharing or data-mapping attributes marked for the Symbols may be used later in the Semantics Analysis and in the Code Generation.

Module File Extensions for OpenMP

After the successful compilation of modules and submodules that may contain the following Declarative Directives, the entire directive starting from !$OMP needs to be written out into .mod files in their corresponding Specification Part:

  • declare simd or declare target

    In the “New Symbol without new Scope” section, we described that when encountering these two declarative directives, new Flag will be applied to the Symbol of the name of the enclosing function, subroutine, or interface body to which it applies, or proc-name. This Flag should be part of the API information for the given subroutine or function

  • declare reduction

    The reduction-identifier in this directive can be use-associated or host-associated. However, it will not act like other Symbols because user may have a reduction name that is the same as a Fortran entity name in the same scope. Therefore a specific data structure needs to be created to save the reduction-identifier information in the Scope and this directive needs to be written into .mod files

Phases of OpenMP Analysis

  1. Create the parse tree for OpenMP
    1. Add types for directives and clauses
      1. Add type(s) that will be used for directives
      2. Add type(s) that will be used for clauses
      3. Add other types, e.g. wrappers or other containers
      4. Use std::variant to encapsulate meaningful types
    2. Implemented in the parser for OpenMP (openmp-grammar.h)
  2. Create canonical nesting
    1. Restructure parse tree to reflect the association of directives and stmts
      1. Associate OpenMPLoopConstruct with DoConstruct and OpenMPEndLoopDirective
    2. Investigate, and perhaps reuse, the algorithm used to restructure do-loops
    3. Add a pass near the code that restructures do-loops; but do not extend the code that handles do-loop for OpenMP; keep this code separate.
    4. Report errors that prevent restructuring (e.g. loop directive not followed by loop) We should abort in case of errors because there is no point to perform further checks if it is not a legal OpenMP construct
  3. Validate the structured-block
    1. Structured-block is a block of executable statements
    2. Single entry and single exit
    3. Access to the structured block must not be the result of a branch
    4. The point of exit cannot be a branch out of the structured block
  4. Check that directive and clause combinations are legal
    1. Begin and End directive should match
    2. Simply check that the clauses are allowed by the directives
    3. Write as a separate pass for simplicity and correctness of the parse tree
  5. Write parse tree tests
    1. At this point, the parse tree should be perfectly formed
    2. Write tests that check for correct form and provenance information
    3. Write tests for errors that can occur during the restructuring
  6. Scope, symbol tables, and name resolution
    1. Update the existing code to handle names and scopes introduced by OpenMP
    2. Write tests to make sure names are properly implemented
  7. Check semantics that is specific to each directive
    1. Validate the directive and its clauses
    2. Some clause checks require the result of name resolution, i.e. “A list item may appear in a linear or firstprivate clause but not both.”
    3. TBD: Validate the nested statement for legality in the scope of the directive
    4. Check the nesting of regions [OpenMP 4.5 spec 2.17]
  8. Module file utilities
    1. Write necessary OpenMP declarative directives to .mod files
    2. Update the existing code to read available OpenMP directives from the .mod files