Added parsing/sema/codegen support for 'parallel master taskloop'
constructs. Some of the clauses, like 'grainsize', 'num_tasks', 'final'
and 'priority' are not supported in full, only constant expressions can
be used currently in these clauses.
llvm-svn: 374791
Added basic support for non-rectangular loops. It requires an additional
analysis of min/max boundaries for non-rectangular loops. Since only
linear dependency is allowed, we can do this analysis.
llvm-svn: 368903
Summary:
https://www.openmp.org/wp-content/uploads/OpenMP-API-Specification-5.0.pdf, page 3:
```
structured block
For C/C++, an executable statement, possibly compound, with a single entry at the
top and a single exit at the bottom, or an OpenMP construct.
COMMENT: See Section 2.1 on page 38 for restrictions on structured
blocks.
```
```
2.1 Directive Format
Some executable directives include a structured block. A structured block:
• may contain infinite loops where the point of exit is never reached;
• may halt due to an IEEE exception;
• may contain calls to exit(), _Exit(), quick_exit(), abort() or functions with a
_Noreturn specifier (in C) or a noreturn attribute (in C/C++);
• may be an expression statement, iteration statement, selection statement, or try block, provided
that the corresponding compound statement obtained by enclosing it in { and } would be a
structured block; and
Restrictions
Restrictions to structured blocks are as follows:
• Entry to a structured block must not be the result of a branch.
• The point of exit cannot be a branch out of the structured block.
C / C++
• The point of entry to a structured block must not be a call to setjmp().
• longjmp() and throw() must not violate the entry/exit criteria.
```
Of particular note here is the fact that OpenMP structured blocks are as-if `noexcept`,
in the same sense as with the normal `noexcept` functions in C++.
I.e. if throw happens, and it attempts to travel out of the `noexcept` function
(here: out of the current structured-block), then the program terminates.
Now, one of course can say that since it is explicitly prohibited by the Specification,
then any and all programs that violate this Specification contain undefined behavior,
and are unspecified, and thus no one should care about them. Just don't write broken code /s
But i'm not sure this is a reasonable approach.
I have personally had oss-fuzz issues of this origin - exception thrown inside
of an OpenMP structured-block that is not caught, thus causing program termination.
This issue isn't all that hard to catch, it's not any particularly different from
diagnosing the same situation with the normal `noexcept` function.
Now, clang static analyzer does not presently model exceptions.
But clang-tidy has a simplisic [[ https://clang.llvm.org/extra/clang-tidy/checks/bugprone-exception-escape.html | bugprone-exception-escape ]] check,
and it is even refactored as a `ExceptionAnalyzer` class for reuse.
So it would be trivial to use that analyzer to check for
exceptions escaping out of OpenMP structured blocks. (D59466)
All that sounds too great to be true. Indeed, there is a caveat.
Presently, it's practically impossible to do. To check a OpenMP structured block
you need to somehow 'get' the OpenMP structured block, and you can't because
it's simply not modelled in AST. `CapturedStmt`/`CapturedDecl` is not it's representation.
Now, it is of course possible to write e.g. some AST matcher that would e.g.
match every OpenMP executable directive, and then return the whatever `Stmt` is
the structured block of said executable directive, if any.
But i said //practically//. This isn't practical for the following reasons:
1. This **will** bitrot. That matcher will need to be kept up-to-date,
and refreshed with every new OpenMP spec version.
2. Every single piece of code that would want that knowledge would need to
have such matcher. Well, okay, if it is an AST matcher, it could be shared.
But then you still have `RecursiveASTVisitor` and friends.
`2 > 1`, so now you have code duplication.
So it would be reasonable (and is fully within clang AST spirit) to not
force every single consumer to do that work, but instead store that knowledge
in the correct, and appropriate place - AST, class structure.
Now, there is another hoop we need to get through.
It isn't fully obvious //how// to model this.
The best solution would of course be to simply add a `OMPStructuredBlock` transparent
node. It would be optimal, it would give us two properties:
* Given this `OMPExecutableDirective`, what's it OpenMP structured block?
* It is trivial to check whether the `Stmt*` is a OpenMP structured block (`isa<OMPStructuredBlock>(ptr)`)
But OpenMP structured block isn't **necessarily** the first, direct child of `OMP*Directive`.
(even ignoring the clang's `CapturedStmt`/`CapturedDecl` that were inserted inbetween).
So i'm not sure whether or not we could re-create AST statements after they were already created?
There would be other costs to a new AST node: https://bugs.llvm.org/show_bug.cgi?id=40563#c12
```
1. You will need to break the representation of loops. The body should be replaced by the "structured block" entity.
2. You will need to support serialization/deserialization.
3. You will need to support template instantiation.
4. You will need to support codegen and take this new construct to account in each OpenMP directive.
```
Instead, there **is** an functionally-equivalent, alternative solution, consisting of two parts.
Part 1:
* Add a member function `isStandaloneDirective()` to the `OMPExecutableDirective` class,
that will tell whether this directive is stand-alone or not, as per the spec.
We need it because we can't just check for the existance of associated statements,
see code comment.
* Add a member function `getStructuredBlock()` to the OMPExecutableDirective` class itself,
that assert that this is not a stand-alone directive, and either return the correct loop body
if this is a loop-like directive, or the captured statement.
This way, given an `OMPExecutableDirective`, we can get it's structured block.
Also, since the knowledge is ingrained into the clang OpenMP implementation,
it will not cause any duplication, and //hopefully// won't bitrot.
Great we achieved 1 of 2 properties of `OMPStructuredBlock` approach.
Thus, there is a second part needed:
* How can we check whether a given `Stmt*` is `OMPStructuredBlock`?
Well, we can't really, in general. I can see this workaround:
```
class FunctionASTVisitor : public RecursiveASTVisitor<FunctionASTVisitor> {
using Base = RecursiveASTVisitor<FunctionASTVisitor>;
public:
bool VisitOMPExecDir(OMPExecDir *D) {
OmpStructuredStmts.emplace_back(D.getStructuredStmt());
}
bool VisitSOMETHINGELSE(???) {
if(InOmpStructuredStmt)
HI!
}
bool TraverseStmt(Stmt *Node) {
if (!Node)
return Base::TraverseStmt(Node);
if (OmpStructuredStmts.back() == Node)
++InOmpStructuredStmt;
Base::TraverseStmt(Node);
if (OmpStructuredStmts.back() == Node) {
OmpStructuredStmts.pop_back();
--InOmpStructuredStmt;
}
return true;
}
std::vector<Stmt*> OmpStructuredStmts;
int InOmpStructuredStmt = 0;
};
```
But i really don't see using it in practice.
It's just too intrusive; and again, requires knowledge duplication.
.. but no. The solution lies right on the ground.
Why don't we simply store this `i'm a openmp structured block` in the bitfield of the `Stmt` itself?
This does not appear to have any impact on the memory footprint of the clang AST,
since it's just a single extra bit in the bitfield. At least the static assertions don't fail.
Thus, indeed, we can achieve both of the properties without a new AST node.
We can cheaply set that bit right in sema, at the end of `Sema::ActOnOpenMPExecutableDirective()`,
by just calling the `getStructuredBlock()` that we just added.
Test coverage that demonstrates all this has been added.
This isn't as great with serialization though. Most of it does not use abbrevs,
so we do end up paying the full price (4 bytes?) instead of a single bit.
That price, of course, can be reclaimed by using abbrevs.
In fact, i suspect that //might// not just reclaim these bytes, but pack these PCH significantly.
I'm not seeing a third solution. If there is one, it would be interesting to hear about it.
("just don't write code that would require `isa<OMPStructuredBlock>(ptr)`" is not a solution.)
Fixes [[ https://bugs.llvm.org/show_bug.cgi?id=40563 | PR40563 ]].
Reviewers: ABataev, rjmccall, hfinkel, rsmith, riccibruno, gribozavr
Reviewed By: ABataev, gribozavr
Subscribers: mgorny, aaron.ballman, steveire, guansong, jfb, jdoerfert, cfe-commits
Tags: #clang, #openmp
Differential Revision: https://reviews.llvm.org/D59214
llvm-svn: 356570
to reflect the new license.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351636
Summary: This patch adds a new code generation path for bound sharing directives containing distribute parallel for. The new code generation scheme applies to chunked schedules on distribute and parallel for directives. The scheme simplifies the code that is being generated by eliminating the need for an outer for loop over chunks for both distribute and parallel for directives. In the case of distribute it applies to any sized chunk while in the parallel for case it only applies when chunk size is 1.
Reviewers: ABataev, caomhin
Reviewed By: ABataev
Subscribers: jholewinski, guansong, cfe-commits
Differential Revision: https://reviews.llvm.org/D53448
llvm-svn: 345509
OpenMP 5.0 introduces asynchronous data update/dependecies clauses on
target data directives. Patch adds initial support for outer task
regions to use task-based codegen for future async target data
directives.
llvm-svn: 318781
https://reviews.llvm.org/D32237
This patch prepares sema with additional fields to support all those composite and combined constructs of OpenMP that include pragma 'distribute' and 'for', such as 'distribute parallel for'. It also extends the regression tests for 'distribute parallel for' and adds a new one.
llvm-svn: 300802
https://reviews.llvm.org/D29922
This patch adds two fields for use in the implementation of 'distribute parallel for':
The increment expression for the distribute loop. As the chunk assigned to a team is executed by multiple threads within the 'parallel for' region, the increment expression has to correspond to the value returned by the related runtime call (for_static_init).
The upper bound of the innermost loop ('for' in 'distribute parallel for') is not the globalUB expression normally used for pragma 'for' when found in isolation. It is instead the upper bound of the chunk assigned to the team ('distribute' loop). In this way, we prevent teams from executing chunks assigned to other teams.
The use of these two fields can be see in a related explanatory patch:
https://reviews.llvm.org/D29508
llvm-svn: 295497
This patch is to implement sema and parsing for 'target teams distribute simd’ pragma.
Differential Revision: https://reviews.llvm.org/D28252
llvm-svn: 291579
This patch is to implement sema and parsing for 'target teams distribute parallel for simd’ pragma.
Differential Revision: https://reviews.llvm.org/D28202
llvm-svn: 290862
This patch is to implement sema and parsing for 'target teams distribute parallel for’ pragma.
Differential Revision: https://reviews.llvm.org/D28160
llvm-svn: 290725
This patch is to implement sema and parsing for 'target teams distribute' pragma.
Differential Revision: https://reviews.llvm.org/D28015
llvm-svn: 290508
This patch is to implement sema and parsing for 'teams distribute parallel for' pragma.
Differential Revision: https://reviews.llvm.org/D27345
llvm-svn: 289179
This patch is to implement sema and parsing for 'teams distribute parallel for simd' pragma.
Differential Revision: https://reviews.llvm.org/D27084
llvm-svn: 288294
This reverts commit r279003 as it breaks some of our buildbots (e.g.
clang-cmake-aarch64-quick, clang-x86_64-linux-selfhost-modules).
The error is in OpenMP/teams_distribute_simd_ast_print.cpp:
clang: /home/buildslave/buildslave/clang-cmake-aarch64-quick/llvm/include/llvm/ADT/DenseMap.h:527:
bool llvm::DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT>::LookupBucketFor(const LookupKeyT&, const BucketT*&) const
[with LookupKeyT = clang::Stmt*; DerivedT = llvm::DenseMap<clang::Stmt*, long unsigned int>;
KeyT = clang::Stmt*; ValueT = long unsigned int;
KeyInfoT = llvm::DenseMapInfo<clang::Stmt*>;
BucketT = llvm::detail::DenseMapPair<clang::Stmt*, long unsigned int>]:
Assertion `!KeyInfoT::isEqual(Val, EmptyKey) && !KeyInfoT::isEqual(Val, TombstoneKey) &&
"Empty/Tombstone value shouldn't be inserted into map!"' failed.
llvm-svn: 279045
This patch is to implement sema and parsing for 'teams distribute simd’ pragma.
This patch is originated by Carlo Bertolli.
Differential Revision: https://reviews.llvm.org/D23528
llvm-svn: 279003
This patch is to implement sema and parsing for 'target parallel for simd' pragma.
Differential Revision: http://reviews.llvm.org/D22096
llvm-svn: 275365
Summary: This patch is an implementation of sema and parsing for the OpenMP composite pragma 'distribute simd'.
Differential Revision: http://reviews.llvm.org/D22007
llvm-svn: 274604
Summary: This patch is an implementation of sema and parsing for the OpenMP composite pragma 'distribute parallel for simd'.
Differential Revision: http://reviews.llvm.org/D21977
llvm-svn: 274530
[OpenMP] Initial implementation of parse and sema for composite pragma 'distribute parallel for'
This patch is an initial implementation for #distribute parallel for.
The main differences that affect other pragmas are:
The implementation of 'distribute parallel for' requires blocking of the associated loop, where blocks are "distributed" to different teams and iterations within each block are scheduled to parallel threads within each team. To implement blocking, sema creates two additional worksharing directive fields that are used to pass the team assigned block lower and upper bounds through the outlined function resulting from 'parallel'. In this way, scheduling for 'for' to threads can use those bounds.
As a consequence of blocking, the stride of 'distribute' is not 1 but it is equal to the blocking size. This is returned by the runtime and sema prepares a DistIncrExpr variable to hold that value.
As a consequence of blocking, the global upper bound (EnsureUpperBound) expression of the 'for' is not the original loop upper bound (e.g. in for(i = 0 ; i < N; i++) this is 'N') but it is the team-assigned block upper bound. Sema creates a new expression holding the calculation of the actual upper bound for 'for' as UB = min(UB, PrevUB), where UB is the loop upper bound, and PrevUB is the team-assigned block upper bound.
llvm-svn: 273884
http://reviews.llvm.org/D21564
This patch is an initial implementation for #distribute parallel for.
The main differences that affect other pragmas are:
The implementation of 'distribute parallel for' requires blocking of the associated loop, where blocks are "distributed" to different teams and iterations within each block are scheduled to parallel threads within each team. To implement blocking, sema creates two additional worksharing directive fields that are used to pass the team assigned block lower and upper bounds through the outlined function resulting from 'parallel'. In this way, scheduling for 'for' to threads can use those bounds.
As a consequence of blocking, the stride of 'distribute' is not 1 but it is equal to the blocking size. This is returned by the runtime and sema prepares a DistIncrExpr variable to hold that value.
As a consequence of blocking, the global upper bound (EnsureUpperBound) expression of the 'for' is not the original loop upper bound (e.g. in for(i = 0 ; i < N; i++) this is 'N') but it is the team-assigned block upper bound. Sema creates a new expression holding the calculation of the actual upper bound for 'for' as UB = min(UB, PrevUB), where UB is the loop upper bound, and PrevUB is the team-assigned block upper bound.
llvm-svn: 273705
Summary:
This patch is to add parsing and sema support for `target update` directive. Support for the `to` and `from` clauses will be added by a different patch. This patch also adds support for other clauses that are already implemented upstream and apply to `target update`, e.g. `device` and `if`.
This patch is based on the original post by Kelvin Li.
Reviewers: hfinkel, carlo.bertolli, kkwli0, arpith-jacob, ABataev
Subscribers: caomhin, cfe-commits
Differential Revision: http://reviews.llvm.org/D15944
llvm-svn: 270878
For better support of some specific GNU extensions some extra
transformation of AST nodes were introduced. These transformations are
very hard to handle. The code is improved in handling of these
extensions by using captured expressions construct.
llvm-svn: 264709
Summary:
This patch adds parsing + sema for the target parallel for directive along with testcases.
Reviewers: ABataev
Differential Revision: http://reviews.llvm.org/D16759
llvm-svn: 259654
Summary:
This patch adds parsing + sema for the target parallel directive and its clauses along with testcases.
Reviewers: ABataev
Differential Revision: http://reviews.llvm.org/D16553
Rebased to current trunk and updated test cases.
llvm-svn: 258832
Alexey Bataev