[OpenACC] Implement loop 'gang' clause. #112006
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
The 'gang' clause is used to specify parallel execution of loops, thus has some complicated rules depending on the 'loop's associated compute construct. This patch implements all of those.
llvmbot
added
clang
|
@llvm/pr-subscribers-clang-modules Author: Erich Keane (erichkeane) ChangesThe 'gang' clause is used to specify parallel execution of loops, thus has some complicated rules depending on the 'loop's associated compute construct. This patch implements all of those. Patch is 75.19 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/112006.diff 22 Files Affected:
diff --git a/clang/include/clang/AST/OpenACCClause.h b/clang/include/clang/AST/OpenACCClause.h
index b500acc768e55a..f3a09eb651458d 100644
--- a/clang/include/clang/AST/OpenACCClause.h
+++ b/clang/include/clang/AST/OpenACCClause.h
@@ -119,32 +119,6 @@ class OpenACCSeqClause : public OpenACCClause {
}
};
-// Not yet implemented, but the type name is necessary for 'seq' diagnostics, so
-// this provides a basic, do-nothing implementation. We still need to add this
-// type to the visitors/etc, as well as get it to take its proper arguments.
-class OpenACCGangClause : public OpenACCClause {
-protected:
- OpenACCGangClause(SourceLocation BeginLoc, SourceLocation EndLoc)
- : OpenACCClause(OpenACCClauseKind::Gang, BeginLoc, EndLoc) {
- llvm_unreachable("Not yet implemented");
- }
-
-public:
- static bool classof(const OpenACCClause *C) {
- return C->getClauseKind() == OpenACCClauseKind::Gang;
- }
-
- static OpenACCGangClause *
- Create(const ASTContext &Ctx, SourceLocation BeginLoc, SourceLocation EndLoc);
-
- child_range children() {
- return child_range(child_iterator(), child_iterator());
- }
- const_child_range children() const {
- return const_child_range(const_child_iterator(), const_child_iterator());
- }
-};
-
// Not yet implemented, but the type name is necessary for 'seq' diagnostics, so
// this provides a basic, do-nothing implementation. We still need to add this
// type to the visitors/etc, as well as get it to take its proper arguments.
@@ -177,7 +151,7 @@ class OpenACCVectorClause : public OpenACCClause {
class OpenACCWorkerClause : public OpenACCClause {
protected:
OpenACCWorkerClause(SourceLocation BeginLoc, SourceLocation EndLoc)
- : OpenACCClause(OpenACCClauseKind::Gang, BeginLoc, EndLoc) {
+ : OpenACCClause(OpenACCClauseKind::Worker, BeginLoc, EndLoc) {
llvm_unreachable("Not yet implemented");
}
@@ -535,6 +509,38 @@ class OpenACCClauseWithSingleIntExpr : public OpenACCClauseWithExprs {
Expr *getIntExpr() { return hasIntExpr() ? getExprs()[0] : nullptr; };
};
+class OpenACCGangClause final
+ : public OpenACCClauseWithExprs,
+ public llvm::TrailingObjects<OpenACCGangClause, Expr *, OpenACCGangKind> {
+protected:
+ OpenACCGangClause(SourceLocation BeginLoc, SourceLocation LParenLoc,
+ ArrayRef<OpenACCGangKind> GangKinds,
+ ArrayRef<Expr *> IntExprs, SourceLocation EndLoc);
+
+ OpenACCGangKind getGangKind(unsigned I) const {
+ return getTrailingObjects<OpenACCGangKind>()[I];
+ }
+
+public:
+ static bool classof(const OpenACCClause *C) {
+ return C->getClauseKind() == OpenACCClauseKind::Gang;
+ }
+
+ size_t numTrailingObjects(OverloadToken<Expr *>) const {
+ return getNumExprs();
+ }
+
+ unsigned getNumExprs() const { return getExprs().size(); }
+ std::pair<OpenACCGangKind, const Expr *> getExpr(unsigned I) const {
+ return {getGangKind(I), getExprs()[I]};
+ }
+
+ static OpenACCGangClause *
+ Create(const ASTContext &Ctx, SourceLocation BeginLoc,
+ SourceLocation LParenLoc, ArrayRef<OpenACCGangKind> GangKinds,
+ ArrayRef<Expr *> IntExprs, SourceLocation EndLoc);
+};
+
class OpenACCNumWorkersClause : public OpenACCClauseWithSingleIntExpr {
OpenACCNumWorkersClause(SourceLocation BeginLoc, SourceLocation LParenLoc,
Expr *IntExpr, SourceLocation EndLoc);
diff --git a/clang/include/clang/Basic/DiagnosticSemaKinds.td b/clang/include/clang/Basic/DiagnosticSemaKinds.td
index 41cdd09e971651..3c62a017005e59 100644
--- a/clang/include/clang/Basic/DiagnosticSemaKinds.td
+++ b/clang/include/clang/Basic/DiagnosticSemaKinds.td
@@ -12576,6 +12576,7 @@ def err_acc_duplicate_clause_disallowed
: Error<"OpenACC '%1' clause cannot appear more than once on a '%0' "
"directive">;
def note_acc_previous_clause_here : Note<"previous clause is here">;
+def note_acc_previous_expr_here : Note<"previous expression is here">;
def err_acc_branch_in_out_compute_construct
: Error<"invalid %select{branch|return|throw}0 %select{out of|into}1 "
"OpenACC Compute Construct">;
@@ -12682,6 +12683,24 @@ def err_acc_insufficient_loops
def err_acc_intervening_code
: Error<"inner loops must be tightly nested inside a '%0' clause on "
"a 'loop' construct">;
+def err_acc_gang_multiple_elt
+ : Error<"OpenACC 'gang' clause may have at most one %select{unnamed or "
+ "'num'|'dim'|'static'}0 argument">;
+def err_acc_gang_arg_invalid
+ : Error<"'%0' argument on 'gang' clause is not permitted on a%select{n "
+ "orphaned|||}1 'loop' construct %select{|associated with a "
+ "'parallel' compute construct|associated with a 'kernels' compute "
+ "construct|associated with a 'serial' compute construct}1">;
+def err_acc_gang_dim_value
+ : Error<"argument to 'gang' clause dimension must be %select{a constant "
+ "expression|1, 2, or 3: evaluated to %1}0">;
+def err_acc_gang_num_gangs_conflict
+ : Error<"'num' argument to 'gang' clause not allowed on a 'loop' construct "
+ "associated with a 'kernels' construct that has a 'num_gangs' "
+ "clause">;
+def err_acc_gang_inside_gang
+ : Error<"loop with a 'gang' clause may not exist in the region of a 'gang' "
+ "clause on a 'kernels' compute construct">;
// AMDGCN builtins diagnostics
def err_amdgcn_global_load_lds_size_invalid_value : Error<"invalid size value">;
diff --git a/clang/include/clang/Basic/OpenACCClauses.def b/clang/include/clang/Basic/OpenACCClauses.def
index a380e5ae69c418..2a098de31eb618 100644
--- a/clang/include/clang/Basic/OpenACCClauses.def
+++ b/clang/include/clang/Basic/OpenACCClauses.def
@@ -42,6 +42,7 @@ VISIT_CLAUSE(DevicePtr)
VISIT_CLAUSE(DeviceType)
CLAUSE_ALIAS(DType, DeviceType, false)
VISIT_CLAUSE(FirstPrivate)
+VISIT_CLAUSE(Gang)
VISIT_CLAUSE(If)
VISIT_CLAUSE(Independent)
VISIT_CLAUSE(NoCreate)
diff --git a/clang/include/clang/Basic/OpenACCKinds.h b/clang/include/clang/Basic/OpenACCKinds.h
index c4dfe3bedc13a7..3f48ebca708a42 100644
--- a/clang/include/clang/Basic/OpenACCKinds.h
+++ b/clang/include/clang/Basic/OpenACCKinds.h
@@ -550,6 +550,35 @@ inline llvm::raw_ostream &operator<<(llvm::raw_ostream &Out,
OpenACCReductionOperator Op) {
return printOpenACCReductionOperator(Out, Op);
}
+
+enum class OpenACCGangKind : uint8_t {
+ /// num:
+ Num,
+ /// dim:
+ Dim,
+ /// static:
+ Static
+};
+
+template <typename StreamTy>
+inline StreamTy &printOpenACCGangKind(StreamTy &Out, OpenACCGangKind GK) {
+ switch (GK) {
+ case OpenACCGangKind::Num:
+ return Out << "num";
+ case OpenACCGangKind::Dim:
+ return Out << "dim";
+ case OpenACCGangKind::Static:
+ return Out << "static";
+ }
+}
+inline const StreamingDiagnostic &operator<<(const StreamingDiagnostic &Out,
+ OpenACCGangKind Op) {
+ return printOpenACCGangKind(Out, Op);
+}
+inline llvm::raw_ostream &operator<<(llvm::raw_ostream &Out,
+ OpenACCGangKind Op) {
+ return printOpenACCGangKind(Out, Op);
+}
} // namespace clang
#endif // LLVM_CLANG_BASIC_OPENACCKINDS_H
diff --git a/clang/include/clang/Parse/Parser.h b/clang/include/clang/Parse/Parser.h
index dbcb545058a026..045ee754a242b3 100644
--- a/clang/include/clang/Parse/Parser.h
+++ b/clang/include/clang/Parse/Parser.h
@@ -3797,9 +3797,15 @@ class Parser : public CodeCompletionHandler {
bool ParseOpenACCSizeExprList(OpenACCClauseKind CK,
llvm::SmallVectorImpl<Expr *> &SizeExprs);
/// Parses a 'gang-arg-list', used for the 'gang' clause.
- bool ParseOpenACCGangArgList(SourceLocation GangLoc);
- /// Parses a 'gang-arg', used for the 'gang' clause.
- bool ParseOpenACCGangArg(SourceLocation GangLoc);
+ bool ParseOpenACCGangArgList(SourceLocation GangLoc,
+ llvm::SmallVectorImpl<OpenACCGangKind> &GKs,
+ llvm::SmallVectorImpl<Expr *> &IntExprs);
+
+ using OpenACCGangArgRes = std::pair<OpenACCGangKind, ExprResult>;
+ /// Parses a 'gang-arg', used for the 'gang' clause. Returns a pair of the
+ /// ExprResult (which contains the validity of the expression), plus the gang
+ /// kind for the current argument.
+ OpenACCGangArgRes ParseOpenACCGangArg(SourceLocation GangLoc);
/// Parses a 'condition' expr, ensuring it results in a
ExprResult ParseOpenACCConditionExpr();
diff --git a/clang/include/clang/Sema/SemaOpenACC.h b/clang/include/clang/Sema/SemaOpenACC.h
index 97386d2378b758..59a9648d5f9380 100644
--- a/clang/include/clang/Sema/SemaOpenACC.h
+++ b/clang/include/clang/Sema/SemaOpenACC.h
@@ -38,9 +38,20 @@ class SemaOpenACC : public SemaBase {
/// haven't had their 'parent' compute construct set yet. Entires will only be
/// made to this list in the case where we know the loop isn't an orphan.
llvm::SmallVector<OpenACCLoopConstruct *> ParentlessLoopConstructs;
- /// Whether we are inside of a compute construct, and should add loops to the
- /// above collection.
- bool InsideComputeConstruct = false;
+
+ struct ComputeConstructInfo {
+ /// Which type of compute construct we are inside of, which we can use to
+ /// determine whether we should add loops to the above collection. We can
+ /// also use it to diagnose loop construct clauses.
+ OpenACCDirectiveKind Kind = OpenACCDirectiveKind::Invalid;
+ // If we have an active compute construct, stores the list of clauses we've
+ // prepared for it, so that we can diagnose limitations on child constructs.
+ ArrayRef<OpenACCClause *> Clauses;
+ } ActiveComputeConstructInfo;
+
+ bool isInComputeConstruct() const {
+ return ActiveComputeConstructInfo.Kind != OpenACCDirectiveKind::Invalid;
+ }
/// Certain clauses care about the same things that aren't specific to the
/// individual clause, but can be shared by a few, so store them here. All
@@ -99,6 +110,15 @@ class SemaOpenACC : public SemaBase {
} TileInfo;
public:
+ ComputeConstructInfo &getActiveComputeConstructInfo() {
+ return ActiveComputeConstructInfo;
+ }
+
+ /// If there is a current 'active' loop construct with a 'gang' clause on a
+ /// 'kernel' construct, this will have the source location for it. This
+ /// permits us to implement the restriction of no further 'gang' clauses.
+ SourceLocation LoopGangClauseOnKernelLoc;
+
// Redeclaration of the version in OpenACCClause.h.
using DeviceTypeArgument = std::pair<IdentifierInfo *, SourceLocation>;
@@ -149,9 +169,14 @@ class SemaOpenACC : public SemaBase {
Expr *LoopCount;
};
+ struct GangDetails {
+ SmallVector<OpenACCGangKind> GangKinds;
+ SmallVector<Expr *> IntExprs;
+ };
+
std::variant<std::monostate, DefaultDetails, ConditionDetails,
IntExprDetails, VarListDetails, WaitDetails, DeviceTypeDetails,
- ReductionDetails, CollapseDetails>
+ ReductionDetails, CollapseDetails, GangDetails>
Details = std::monostate{};
public:
@@ -245,9 +270,18 @@ class SemaOpenACC : public SemaBase {
ClauseKind == OpenACCClauseKind::NumWorkers ||
ClauseKind == OpenACCClauseKind::Async ||
ClauseKind == OpenACCClauseKind::Tile ||
+ ClauseKind == OpenACCClauseKind::Gang ||
ClauseKind == OpenACCClauseKind::VectorLength) &&
"Parsed clause kind does not have a int exprs");
+ if (ClauseKind == OpenACCClauseKind::Gang) {
+ // There might not be any gang int exprs, as this is an optional
+ // argument.
+ if (std::holds_alternative<std::monostate>(Details))
+ return {};
+ return std::get<GangDetails>(Details).IntExprs;
+ }
+
return std::get<IntExprDetails>(Details).IntExprs;
}
@@ -259,6 +293,16 @@ class SemaOpenACC : public SemaBase {
return std::get<ReductionDetails>(Details).Op;
}
+ ArrayRef<OpenACCGangKind> getGangKinds() const {
+ assert(ClauseKind == OpenACCClauseKind::Gang &&
+ "Parsed clause kind does not have gang kind");
+ // The args on gang are optional, so this might not actually hold
+ // anything.
+ if (std::holds_alternative<std::monostate>(Details))
+ return {};
+ return std::get<GangDetails>(Details).GangKinds;
+ }
+
ArrayRef<Expr *> getVarList() {
assert((ClauseKind == OpenACCClauseKind::Private ||
ClauseKind == OpenACCClauseKind::NoCreate ||
@@ -371,6 +415,25 @@ class SemaOpenACC : public SemaBase {
Details = IntExprDetails{std::move(IntExprs)};
}
+ void setGangDetails(ArrayRef<OpenACCGangKind> GKs,
+ ArrayRef<Expr *> IntExprs) {
+ assert(ClauseKind == OpenACCClauseKind::Gang &&
+ "Parsed Clause kind does not have gang details");
+ assert(GKs.size() == IntExprs.size() && "Mismatched kind/size?");
+
+ Details = GangDetails{{GKs.begin(), GKs.end()},
+ {IntExprs.begin(), IntExprs.end()}};
+ }
+
+ void setGangDetails(llvm::SmallVector<OpenACCGangKind> &&GKs,
+ llvm::SmallVector<Expr *> &&IntExprs) {
+ assert(ClauseKind == OpenACCClauseKind::Gang &&
+ "Parsed Clause kind does not have gang details");
+ assert(GKs.size() == IntExprs.size() && "Mismatched kind/size?");
+
+ Details = GangDetails{std::move(GKs), std::move(IntExprs)};
+ }
+
void setVarListDetails(ArrayRef<Expr *> VarList, bool IsReadOnly,
bool IsZero) {
assert((ClauseKind == OpenACCClauseKind::Private ||
@@ -545,10 +608,12 @@ class SemaOpenACC : public SemaBase {
SourceLocation RBLoc);
/// Checks the loop depth value for a collapse clause.
ExprResult CheckCollapseLoopCount(Expr *LoopCount);
- /// Checks a single size expr for a tile clause. 'gang' could possibly call
- /// this, but has slightly stricter rules as to valid values.
+ /// Checks a single size expr for a tile clause.
ExprResult CheckTileSizeExpr(Expr *SizeExpr);
+ // Check a single expression on a gang clause.
+ ExprResult CheckGangExpr(OpenACCGangKind GK, Expr *E);
+
ExprResult BuildOpenACCAsteriskSizeExpr(SourceLocation AsteriskLoc);
ExprResult ActOnOpenACCAsteriskSizeExpr(SourceLocation AsteriskLoc);
@@ -595,8 +660,9 @@ class SemaOpenACC : public SemaBase {
/// Loop needing its parent construct.
class AssociatedStmtRAII {
SemaOpenACC &SemaRef;
- bool WasInsideComputeConstruct;
+ ComputeConstructInfo OldActiveComputeConstructInfo;
OpenACCDirectiveKind DirKind;
+ SourceLocation OldLoopGangClauseOnKernelLoc;
llvm::SmallVector<OpenACCLoopConstruct *> ParentlessLoopConstructs;
LoopInConstructRAII LoopRAII;
diff --git a/clang/lib/AST/OpenACCClause.cpp b/clang/lib/AST/OpenACCClause.cpp
index 0b34ed6189593e..6fb8fe0b8cfeef 100644
--- a/clang/lib/AST/OpenACCClause.cpp
+++ b/clang/lib/AST/OpenACCClause.cpp
@@ -26,7 +26,7 @@ bool OpenACCClauseWithExprs::classof(const OpenACCClause *C) {
return OpenACCWaitClause::classof(C) || OpenACCNumGangsClause::classof(C) ||
OpenACCTileClause::classof(C) ||
OpenACCClauseWithSingleIntExpr::classof(C) ||
- OpenACCClauseWithVarList::classof(C);
+ OpenACCGangClause::classof(C) || OpenACCClauseWithVarList::classof(C);
}
bool OpenACCClauseWithVarList::classof(const OpenACCClause *C) {
return OpenACCPrivateClause::classof(C) ||
@@ -125,6 +125,21 @@ OpenACCNumWorkersClause::OpenACCNumWorkersClause(SourceLocation BeginLoc,
"Condition expression type not scalar/dependent");
}
+OpenACCGangClause::OpenACCGangClause(SourceLocation BeginLoc,
+ SourceLocation LParenLoc,
+ ArrayRef<OpenACCGangKind> GangKinds,
+ ArrayRef<Expr *> IntExprs,
+ SourceLocation EndLoc)
+ : OpenACCClauseWithExprs(OpenACCClauseKind::Gang, BeginLoc, LParenLoc,
+ EndLoc) {
+ assert(GangKinds.size() == IntExprs.size() && "Mismatch exprs/kind?");
+ std::uninitialized_copy(IntExprs.begin(), IntExprs.end(),
+ getTrailingObjects<Expr *>());
+ setExprs(MutableArrayRef(getTrailingObjects<Expr *>(), IntExprs.size()));
+ std::uninitialized_copy(GangKinds.begin(), GangKinds.end(),
+ getTrailingObjects<OpenACCGangKind>());
+}
+
OpenACCNumWorkersClause *
OpenACCNumWorkersClause::Create(const ASTContext &C, SourceLocation BeginLoc,
SourceLocation LParenLoc, Expr *IntExpr,
@@ -376,11 +391,16 @@ OpenACCSeqClause *OpenACCSeqClause::Create(const ASTContext &C,
return new (Mem) OpenACCSeqClause(BeginLoc, EndLoc);
}
-OpenACCGangClause *OpenACCGangClause::Create(const ASTContext &C,
- SourceLocation BeginLoc,
- SourceLocation EndLoc) {
- void *Mem = C.Allocate(sizeof(OpenACCGangClause));
- return new (Mem) OpenACCGangClause(BeginLoc, EndLoc);
+OpenACCGangClause *
+OpenACCGangClause::Create(const ASTContext &C, SourceLocation BeginLoc,
+ SourceLocation LParenLoc,
+ ArrayRef<OpenACCGangKind> GangKinds,
+ ArrayRef<Expr *> IntExprs, SourceLocation EndLoc) {
+ void *Mem =
+ C.Allocate(OpenACCGangClause::totalSizeToAlloc<Expr *, OpenACCGangKind>(
+ IntExprs.size(), GangKinds.size()));
+ return new (Mem)
+ OpenACCGangClause(BeginLoc, LParenLoc, GangKinds, IntExprs, EndLoc);
}
OpenACCWorkerClause *OpenACCWorkerClause::Create(const ASTContext &C,
@@ -600,3 +620,21 @@ void OpenACCClausePrinter::VisitCollapseClause(const OpenACCCollapseClause &C) {
printExpr(C.getLoopCount());
OS << ")";
}
+
+void OpenACCClausePrinter::VisitGangClause(const OpenACCGangClause &C) {
+ OS << "gang";
+
+ if (C.getNumExprs() > 0) {
+ OS << "(";
+ bool first = true;
+ for (unsigned I = 0; I < C.getNumExprs(); ++I) {
+ if (!first)
+ OS << ", ";
+ first = false;
+
+ OS << C.getExpr(I).first << ": ";
+ printExpr(C.getExpr(I).second);
+ }
+ OS << ")";
+ }
+}
diff --git a/clang/lib/AST/StmtProfile.cpp b/clang/lib/AST/StmtProfile.cpp
index 4d177fd6c5968c..6161b1403ed35d 100644
--- a/clang/lib/AST/StmtProfile.cpp
+++ b/clang/lib/AST/StmtProfile.cpp
@@ -2646,6 +2646,12 @@ void OpenACCClauseProfiler::VisitIndependentClause(
void OpenACCClauseProfiler::VisitSeqClause(const OpenACCSeqClause &Clause) {}
+void OpenACCClauseProfiler::VisitGangClause(const OpenACCGangClause &Clause) {
+ for (unsigned I = 0; I < Clause.getNumExprs(); ++I) {
+ Profiler.VisitStmt(Clause.getExpr(I).second);
+ }
+}
+
void OpenACCClauseProfiler::VisitReductionClause(
const OpenACCReductionClause &Clause) {
for (auto *E : Clause.getVarList())
diff --git a/clang/lib/AST/TextNodeDumper.cpp b/clang/lib/AST/TextNodeDumper.cpp
index 15b23d60c3ffab..ac8c196777f9b8 100644
--- a/clang/lib/AST/TextNodeDumper.cpp
+++ b/clang/lib/AST/TextNodeDumper.cpp
@@ -425,6 +425,17 @@ void TextNodeDumper::Visit(const OpenACCClause *C) {
// but print 'clause' here so it is clear what is happening from the dump.
OS << " clause";
break;
+ case OpenACCClauseKind::Gang: {
+ OS << " clause";
+ // print the list of all GangKinds, so that there is some sort of
+ // relationship to the expressions listed afterwards.
+ auto *GC = cast<OpenACCGangClause>(C);
+
+ for (unsigned I = 0; I < GC->getNumExprs(); ++I) {
+ OS << " " << GC->getExpr(I).first;
+ }
+ break;
+ }
case OpenACCClauseKind::Collapse:
OS << " clause";
if (cast<OpenACCCollapseClause>(C)->hasForce())
diff --git a/clang/lib/Parse/ParseOpenACC.cpp b/clang/lib/Parse/ParseOpenACC.cpp
index b27e50b147f4a8..635039b724e6a0 100644
--- a/clang/lib/Parse/ParseOpenACC.cpp
+++ b/clang/lib/Parse/ParseOpenACC.cpp
@@ -797,23 +797,26 @@ bool Parser::ParseOpenACCSizeExprList(
/// [num:]int-expr
/// dim:int-expr
/// static:size-expr
-bool Parser::ParseOpenACCGangArg(SourceLocation GangLoc) {
+Parser::OpenACCGangArgRes Parser::ParseOpenACCGangArg(SourceL...
[truncated]
|
|
@llvm/pr-subscribers-clang Author: Erich Keane (erichkeane) ChangesThe 'gang' clause is used to specify parallel execution of loops, thus has some complicated rules depending on the 'loop's associated compute construct. This patch implements all of those. Patch is 75.19 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/112006.diff 22 Files Affected:
diff --git a/clang/include/clang/AST/OpenACCClause.h b/clang/include/clang/AST/OpenACCClause.h
index b500acc768e55a..f3a09eb651458d 100644
--- a/clang/include/clang/AST/OpenACCClause.h
+++ b/clang/include/clang/AST/OpenACCClause.h
@@ -119,32 +119,6 @@ class OpenACCSeqClause : public OpenACCClause {
}
};
-// Not yet implemented, but the type name is necessary for 'seq' diagnostics, so
-// this provides a basic, do-nothing implementation. We still need to add this
-// type to the visitors/etc, as well as get it to take its proper arguments.
-class OpenACCGangClause : public OpenACCClause {
-protected:
- OpenACCGangClause(SourceLocation BeginLoc, SourceLocation EndLoc)
- : OpenACCClause(OpenACCClauseKind::Gang, BeginLoc, EndLoc) {
- llvm_unreachable("Not yet implemented");
- }
-
-public:
- static bool classof(const OpenACCClause *C) {
- return C->getClauseKind() == OpenACCClauseKind::Gang;
- }
-
- static OpenACCGangClause *
- Create(const ASTContext &Ctx, SourceLocation BeginLoc, SourceLocation EndLoc);
-
- child_range children() {
- return child_range(child_iterator(), child_iterator());
- }
- const_child_range children() const {
- return const_child_range(const_child_iterator(), const_child_iterator());
- }
-};
-
// Not yet implemented, but the type name is necessary for 'seq' diagnostics, so
// this provides a basic, do-nothing implementation. We still need to add this
// type to the visitors/etc, as well as get it to take its proper arguments.
@@ -177,7 +151,7 @@ class OpenACCVectorClause : public OpenACCClause {
class OpenACCWorkerClause : public OpenACCClause {
protected:
OpenACCWorkerClause(SourceLocation BeginLoc, SourceLocation EndLoc)
- : OpenACCClause(OpenACCClauseKind::Gang, BeginLoc, EndLoc) {
+ : OpenACCClause(OpenACCClauseKind::Worker, BeginLoc, EndLoc) {
llvm_unreachable("Not yet implemented");
}
@@ -535,6 +509,38 @@ class OpenACCClauseWithSingleIntExpr : public OpenACCClauseWithExprs {
Expr *getIntExpr() { return hasIntExpr() ? getExprs()[0] : nullptr; };
};
+class OpenACCGangClause final
+ : public OpenACCClauseWithExprs,
+ public llvm::TrailingObjects<OpenACCGangClause, Expr *, OpenACCGangKind> {
+protected:
+ OpenACCGangClause(SourceLocation BeginLoc, SourceLocation LParenLoc,
+ ArrayRef<OpenACCGangKind> GangKinds,
+ ArrayRef<Expr *> IntExprs, SourceLocation EndLoc);
+
+ OpenACCGangKind getGangKind(unsigned I) const {
+ return getTrailingObjects<OpenACCGangKind>()[I];
+ }
+
+public:
+ static bool classof(const OpenACCClause *C) {
+ return C->getClauseKind() == OpenACCClauseKind::Gang;
+ }
+
+ size_t numTrailingObjects(OverloadToken<Expr *>) const {
+ return getNumExprs();
+ }
+
+ unsigned getNumExprs() const { return getExprs().size(); }
+ std::pair<OpenACCGangKind, const Expr *> getExpr(unsigned I) const {
+ return {getGangKind(I), getExprs()[I]};
+ }
+
+ static OpenACCGangClause *
+ Create(const ASTContext &Ctx, SourceLocation BeginLoc,
+ SourceLocation LParenLoc, ArrayRef<OpenACCGangKind> GangKinds,
+ ArrayRef<Expr *> IntExprs, SourceLocation EndLoc);
+};
+
class OpenACCNumWorkersClause : public OpenACCClauseWithSingleIntExpr {
OpenACCNumWorkersClause(SourceLocation BeginLoc, SourceLocation LParenLoc,
Expr *IntExpr, SourceLocation EndLoc);
diff --git a/clang/include/clang/Basic/DiagnosticSemaKinds.td b/clang/include/clang/Basic/DiagnosticSemaKinds.td
index 41cdd09e971651..3c62a017005e59 100644
--- a/clang/include/clang/Basic/DiagnosticSemaKinds.td
+++ b/clang/include/clang/Basic/DiagnosticSemaKinds.td
@@ -12576,6 +12576,7 @@ def err_acc_duplicate_clause_disallowed
: Error<"OpenACC '%1' clause cannot appear more than once on a '%0' "
"directive">;
def note_acc_previous_clause_here : Note<"previous clause is here">;
+def note_acc_previous_expr_here : Note<"previous expression is here">;
def err_acc_branch_in_out_compute_construct
: Error<"invalid %select{branch|return|throw}0 %select{out of|into}1 "
"OpenACC Compute Construct">;
@@ -12682,6 +12683,24 @@ def err_acc_insufficient_loops
def err_acc_intervening_code
: Error<"inner loops must be tightly nested inside a '%0' clause on "
"a 'loop' construct">;
+def err_acc_gang_multiple_elt
+ : Error<"OpenACC 'gang' clause may have at most one %select{unnamed or "
+ "'num'|'dim'|'static'}0 argument">;
+def err_acc_gang_arg_invalid
+ : Error<"'%0' argument on 'gang' clause is not permitted on a%select{n "
+ "orphaned|||}1 'loop' construct %select{|associated with a "
+ "'parallel' compute construct|associated with a 'kernels' compute "
+ "construct|associated with a 'serial' compute construct}1">;
+def err_acc_gang_dim_value
+ : Error<"argument to 'gang' clause dimension must be %select{a constant "
+ "expression|1, 2, or 3: evaluated to %1}0">;
+def err_acc_gang_num_gangs_conflict
+ : Error<"'num' argument to 'gang' clause not allowed on a 'loop' construct "
+ "associated with a 'kernels' construct that has a 'num_gangs' "
+ "clause">;
+def err_acc_gang_inside_gang
+ : Error<"loop with a 'gang' clause may not exist in the region of a 'gang' "
+ "clause on a 'kernels' compute construct">;
// AMDGCN builtins diagnostics
def err_amdgcn_global_load_lds_size_invalid_value : Error<"invalid size value">;
diff --git a/clang/include/clang/Basic/OpenACCClauses.def b/clang/include/clang/Basic/OpenACCClauses.def
index a380e5ae69c418..2a098de31eb618 100644
--- a/clang/include/clang/Basic/OpenACCClauses.def
+++ b/clang/include/clang/Basic/OpenACCClauses.def
@@ -42,6 +42,7 @@ VISIT_CLAUSE(DevicePtr)
VISIT_CLAUSE(DeviceType)
CLAUSE_ALIAS(DType, DeviceType, false)
VISIT_CLAUSE(FirstPrivate)
+VISIT_CLAUSE(Gang)
VISIT_CLAUSE(If)
VISIT_CLAUSE(Independent)
VISIT_CLAUSE(NoCreate)
diff --git a/clang/include/clang/Basic/OpenACCKinds.h b/clang/include/clang/Basic/OpenACCKinds.h
index c4dfe3bedc13a7..3f48ebca708a42 100644
--- a/clang/include/clang/Basic/OpenACCKinds.h
+++ b/clang/include/clang/Basic/OpenACCKinds.h
@@ -550,6 +550,35 @@ inline llvm::raw_ostream &operator<<(llvm::raw_ostream &Out,
OpenACCReductionOperator Op) {
return printOpenACCReductionOperator(Out, Op);
}
+
+enum class OpenACCGangKind : uint8_t {
+ /// num:
+ Num,
+ /// dim:
+ Dim,
+ /// static:
+ Static
+};
+
+template <typename StreamTy>
+inline StreamTy &printOpenACCGangKind(StreamTy &Out, OpenACCGangKind GK) {
+ switch (GK) {
+ case OpenACCGangKind::Num:
+ return Out << "num";
+ case OpenACCGangKind::Dim:
+ return Out << "dim";
+ case OpenACCGangKind::Static:
+ return Out << "static";
+ }
+}
+inline const StreamingDiagnostic &operator<<(const StreamingDiagnostic &Out,
+ OpenACCGangKind Op) {
+ return printOpenACCGangKind(Out, Op);
+}
+inline llvm::raw_ostream &operator<<(llvm::raw_ostream &Out,
+ OpenACCGangKind Op) {
+ return printOpenACCGangKind(Out, Op);
+}
} // namespace clang
#endif // LLVM_CLANG_BASIC_OPENACCKINDS_H
diff --git a/clang/include/clang/Parse/Parser.h b/clang/include/clang/Parse/Parser.h
index dbcb545058a026..045ee754a242b3 100644
--- a/clang/include/clang/Parse/Parser.h
+++ b/clang/include/clang/Parse/Parser.h
@@ -3797,9 +3797,15 @@ class Parser : public CodeCompletionHandler {
bool ParseOpenACCSizeExprList(OpenACCClauseKind CK,
llvm::SmallVectorImpl<Expr *> &SizeExprs);
/// Parses a 'gang-arg-list', used for the 'gang' clause.
- bool ParseOpenACCGangArgList(SourceLocation GangLoc);
- /// Parses a 'gang-arg', used for the 'gang' clause.
- bool ParseOpenACCGangArg(SourceLocation GangLoc);
+ bool ParseOpenACCGangArgList(SourceLocation GangLoc,
+ llvm::SmallVectorImpl<OpenACCGangKind> &GKs,
+ llvm::SmallVectorImpl<Expr *> &IntExprs);
+
+ using OpenACCGangArgRes = std::pair<OpenACCGangKind, ExprResult>;
+ /// Parses a 'gang-arg', used for the 'gang' clause. Returns a pair of the
+ /// ExprResult (which contains the validity of the expression), plus the gang
+ /// kind for the current argument.
+ OpenACCGangArgRes ParseOpenACCGangArg(SourceLocation GangLoc);
/// Parses a 'condition' expr, ensuring it results in a
ExprResult ParseOpenACCConditionExpr();
diff --git a/clang/include/clang/Sema/SemaOpenACC.h b/clang/include/clang/Sema/SemaOpenACC.h
index 97386d2378b758..59a9648d5f9380 100644
--- a/clang/include/clang/Sema/SemaOpenACC.h
+++ b/clang/include/clang/Sema/SemaOpenACC.h
@@ -38,9 +38,20 @@ class SemaOpenACC : public SemaBase {
/// haven't had their 'parent' compute construct set yet. Entires will only be
/// made to this list in the case where we know the loop isn't an orphan.
llvm::SmallVector<OpenACCLoopConstruct *> ParentlessLoopConstructs;
- /// Whether we are inside of a compute construct, and should add loops to the
- /// above collection.
- bool InsideComputeConstruct = false;
+
+ struct ComputeConstructInfo {
+ /// Which type of compute construct we are inside of, which we can use to
+ /// determine whether we should add loops to the above collection. We can
+ /// also use it to diagnose loop construct clauses.
+ OpenACCDirectiveKind Kind = OpenACCDirectiveKind::Invalid;
+ // If we have an active compute construct, stores the list of clauses we've
+ // prepared for it, so that we can diagnose limitations on child constructs.
+ ArrayRef<OpenACCClause *> Clauses;
+ } ActiveComputeConstructInfo;
+
+ bool isInComputeConstruct() const {
+ return ActiveComputeConstructInfo.Kind != OpenACCDirectiveKind::Invalid;
+ }
/// Certain clauses care about the same things that aren't specific to the
/// individual clause, but can be shared by a few, so store them here. All
@@ -99,6 +110,15 @@ class SemaOpenACC : public SemaBase {
} TileInfo;
public:
+ ComputeConstructInfo &getActiveComputeConstructInfo() {
+ return ActiveComputeConstructInfo;
+ }
+
+ /// If there is a current 'active' loop construct with a 'gang' clause on a
+ /// 'kernel' construct, this will have the source location for it. This
+ /// permits us to implement the restriction of no further 'gang' clauses.
+ SourceLocation LoopGangClauseOnKernelLoc;
+
// Redeclaration of the version in OpenACCClause.h.
using DeviceTypeArgument = std::pair<IdentifierInfo *, SourceLocation>;
@@ -149,9 +169,14 @@ class SemaOpenACC : public SemaBase {
Expr *LoopCount;
};
+ struct GangDetails {
+ SmallVector<OpenACCGangKind> GangKinds;
+ SmallVector<Expr *> IntExprs;
+ };
+
std::variant<std::monostate, DefaultDetails, ConditionDetails,
IntExprDetails, VarListDetails, WaitDetails, DeviceTypeDetails,
- ReductionDetails, CollapseDetails>
+ ReductionDetails, CollapseDetails, GangDetails>
Details = std::monostate{};
public:
@@ -245,9 +270,18 @@ class SemaOpenACC : public SemaBase {
ClauseKind == OpenACCClauseKind::NumWorkers ||
ClauseKind == OpenACCClauseKind::Async ||
ClauseKind == OpenACCClauseKind::Tile ||
+ ClauseKind == OpenACCClauseKind::Gang ||
ClauseKind == OpenACCClauseKind::VectorLength) &&
"Parsed clause kind does not have a int exprs");
+ if (ClauseKind == OpenACCClauseKind::Gang) {
+ // There might not be any gang int exprs, as this is an optional
+ // argument.
+ if (std::holds_alternative<std::monostate>(Details))
+ return {};
+ return std::get<GangDetails>(Details).IntExprs;
+ }
+
return std::get<IntExprDetails>(Details).IntExprs;
}
@@ -259,6 +293,16 @@ class SemaOpenACC : public SemaBase {
return std::get<ReductionDetails>(Details).Op;
}
+ ArrayRef<OpenACCGangKind> getGangKinds() const {
+ assert(ClauseKind == OpenACCClauseKind::Gang &&
+ "Parsed clause kind does not have gang kind");
+ // The args on gang are optional, so this might not actually hold
+ // anything.
+ if (std::holds_alternative<std::monostate>(Details))
+ return {};
+ return std::get<GangDetails>(Details).GangKinds;
+ }
+
ArrayRef<Expr *> getVarList() {
assert((ClauseKind == OpenACCClauseKind::Private ||
ClauseKind == OpenACCClauseKind::NoCreate ||
@@ -371,6 +415,25 @@ class SemaOpenACC : public SemaBase {
Details = IntExprDetails{std::move(IntExprs)};
}
+ void setGangDetails(ArrayRef<OpenACCGangKind> GKs,
+ ArrayRef<Expr *> IntExprs) {
+ assert(ClauseKind == OpenACCClauseKind::Gang &&
+ "Parsed Clause kind does not have gang details");
+ assert(GKs.size() == IntExprs.size() && "Mismatched kind/size?");
+
+ Details = GangDetails{{GKs.begin(), GKs.end()},
+ {IntExprs.begin(), IntExprs.end()}};
+ }
+
+ void setGangDetails(llvm::SmallVector<OpenACCGangKind> &&GKs,
+ llvm::SmallVector<Expr *> &&IntExprs) {
+ assert(ClauseKind == OpenACCClauseKind::Gang &&
+ "Parsed Clause kind does not have gang details");
+ assert(GKs.size() == IntExprs.size() && "Mismatched kind/size?");
+
+ Details = GangDetails{std::move(GKs), std::move(IntExprs)};
+ }
+
void setVarListDetails(ArrayRef<Expr *> VarList, bool IsReadOnly,
bool IsZero) {
assert((ClauseKind == OpenACCClauseKind::Private ||
@@ -545,10 +608,12 @@ class SemaOpenACC : public SemaBase {
SourceLocation RBLoc);
/// Checks the loop depth value for a collapse clause.
ExprResult CheckCollapseLoopCount(Expr *LoopCount);
- /// Checks a single size expr for a tile clause. 'gang' could possibly call
- /// this, but has slightly stricter rules as to valid values.
+ /// Checks a single size expr for a tile clause.
ExprResult CheckTileSizeExpr(Expr *SizeExpr);
+ // Check a single expression on a gang clause.
+ ExprResult CheckGangExpr(OpenACCGangKind GK, Expr *E);
+
ExprResult BuildOpenACCAsteriskSizeExpr(SourceLocation AsteriskLoc);
ExprResult ActOnOpenACCAsteriskSizeExpr(SourceLocation AsteriskLoc);
@@ -595,8 +660,9 @@ class SemaOpenACC : public SemaBase {
/// Loop needing its parent construct.
class AssociatedStmtRAII {
SemaOpenACC &SemaRef;
- bool WasInsideComputeConstruct;
+ ComputeConstructInfo OldActiveComputeConstructInfo;
OpenACCDirectiveKind DirKind;
+ SourceLocation OldLoopGangClauseOnKernelLoc;
llvm::SmallVector<OpenACCLoopConstruct *> ParentlessLoopConstructs;
LoopInConstructRAII LoopRAII;
diff --git a/clang/lib/AST/OpenACCClause.cpp b/clang/lib/AST/OpenACCClause.cpp
index 0b34ed6189593e..6fb8fe0b8cfeef 100644
--- a/clang/lib/AST/OpenACCClause.cpp
+++ b/clang/lib/AST/OpenACCClause.cpp
@@ -26,7 +26,7 @@ bool OpenACCClauseWithExprs::classof(const OpenACCClause *C) {
return OpenACCWaitClause::classof(C) || OpenACCNumGangsClause::classof(C) ||
OpenACCTileClause::classof(C) ||
OpenACCClauseWithSingleIntExpr::classof(C) ||
- OpenACCClauseWithVarList::classof(C);
+ OpenACCGangClause::classof(C) || OpenACCClauseWithVarList::classof(C);
}
bool OpenACCClauseWithVarList::classof(const OpenACCClause *C) {
return OpenACCPrivateClause::classof(C) ||
@@ -125,6 +125,21 @@ OpenACCNumWorkersClause::OpenACCNumWorkersClause(SourceLocation BeginLoc,
"Condition expression type not scalar/dependent");
}
+OpenACCGangClause::OpenACCGangClause(SourceLocation BeginLoc,
+ SourceLocation LParenLoc,
+ ArrayRef<OpenACCGangKind> GangKinds,
+ ArrayRef<Expr *> IntExprs,
+ SourceLocation EndLoc)
+ : OpenACCClauseWithExprs(OpenACCClauseKind::Gang, BeginLoc, LParenLoc,
+ EndLoc) {
+ assert(GangKinds.size() == IntExprs.size() && "Mismatch exprs/kind?");
+ std::uninitialized_copy(IntExprs.begin(), IntExprs.end(),
+ getTrailingObjects<Expr *>());
+ setExprs(MutableArrayRef(getTrailingObjects<Expr *>(), IntExprs.size()));
+ std::uninitialized_copy(GangKinds.begin(), GangKinds.end(),
+ getTrailingObjects<OpenACCGangKind>());
+}
+
OpenACCNumWorkersClause *
OpenACCNumWorkersClause::Create(const ASTContext &C, SourceLocation BeginLoc,
SourceLocation LParenLoc, Expr *IntExpr,
@@ -376,11 +391,16 @@ OpenACCSeqClause *OpenACCSeqClause::Create(const ASTContext &C,
return new (Mem) OpenACCSeqClause(BeginLoc, EndLoc);
}
-OpenACCGangClause *OpenACCGangClause::Create(const ASTContext &C,
- SourceLocation BeginLoc,
- SourceLocation EndLoc) {
- void *Mem = C.Allocate(sizeof(OpenACCGangClause));
- return new (Mem) OpenACCGangClause(BeginLoc, EndLoc);
+OpenACCGangClause *
+OpenACCGangClause::Create(const ASTContext &C, SourceLocation BeginLoc,
+ SourceLocation LParenLoc,
+ ArrayRef<OpenACCGangKind> GangKinds,
+ ArrayRef<Expr *> IntExprs, SourceLocation EndLoc) {
+ void *Mem =
+ C.Allocate(OpenACCGangClause::totalSizeToAlloc<Expr *, OpenACCGangKind>(
+ IntExprs.size(), GangKinds.size()));
+ return new (Mem)
+ OpenACCGangClause(BeginLoc, LParenLoc, GangKinds, IntExprs, EndLoc);
}
OpenACCWorkerClause *OpenACCWorkerClause::Create(const ASTContext &C,
@@ -600,3 +620,21 @@ void OpenACCClausePrinter::VisitCollapseClause(const OpenACCCollapseClause &C) {
printExpr(C.getLoopCount());
OS << ")";
}
+
+void OpenACCClausePrinter::VisitGangClause(const OpenACCGangClause &C) {
+ OS << "gang";
+
+ if (C.getNumExprs() > 0) {
+ OS << "(";
+ bool first = true;
+ for (unsigned I = 0; I < C.getNumExprs(); ++I) {
+ if (!first)
+ OS << ", ";
+ first = false;
+
+ OS << C.getExpr(I).first << ": ";
+ printExpr(C.getExpr(I).second);
+ }
+ OS << ")";
+ }
+}
diff --git a/clang/lib/AST/StmtProfile.cpp b/clang/lib/AST/StmtProfile.cpp
index 4d177fd6c5968c..6161b1403ed35d 100644
--- a/clang/lib/AST/StmtProfile.cpp
+++ b/clang/lib/AST/StmtProfile.cpp
@@ -2646,6 +2646,12 @@ void OpenACCClauseProfiler::VisitIndependentClause(
void OpenACCClauseProfiler::VisitSeqClause(const OpenACCSeqClause &Clause) {}
+void OpenACCClauseProfiler::VisitGangClause(const OpenACCGangClause &Clause) {
+ for (unsigned I = 0; I < Clause.getNumExprs(); ++I) {
+ Profiler.VisitStmt(Clause.getExpr(I).second);
+ }
+}
+
void OpenACCClauseProfiler::VisitReductionClause(
const OpenACCReductionClause &Clause) {
for (auto *E : Clause.getVarList())
diff --git a/clang/lib/AST/TextNodeDumper.cpp b/clang/lib/AST/TextNodeDumper.cpp
index 15b23d60c3ffab..ac8c196777f9b8 100644
--- a/clang/lib/AST/TextNodeDumper.cpp
+++ b/clang/lib/AST/TextNodeDumper.cpp
@@ -425,6 +425,17 @@ void TextNodeDumper::Visit(const OpenACCClause *C) {
// but print 'clause' here so it is clear what is happening from the dump.
OS << " clause";
break;
+ case OpenACCClauseKind::Gang: {
+ OS << " clause";
+ // print the list of all GangKinds, so that there is some sort of
+ // relationship to the expressions listed afterwards.
+ auto *GC = cast<OpenACCGangClause>(C);
+
+ for (unsigned I = 0; I < GC->getNumExprs(); ++I) {
+ OS << " " << GC->getExpr(I).first;
+ }
+ break;
+ }
case OpenACCClauseKind::Collapse:
OS << " clause";
if (cast<OpenACCCollapseClause>(C)->hasForce())
diff --git a/clang/lib/Parse/ParseOpenACC.cpp b/clang/lib/Parse/ParseOpenACC.cpp
index b27e50b147f4a8..635039b724e6a0 100644
--- a/clang/lib/Parse/ParseOpenACC.cpp
+++ b/clang/lib/Parse/ParseOpenACC.cpp
@@ -797,23 +797,26 @@ bool Parser::ParseOpenACCSizeExprList(
/// [num:]int-expr
/// dim:int-expr
/// static:size-expr
-bool Parser::ParseOpenACCGangArg(SourceLocation GangLoc) {
+Parser::OpenACCGangArgRes Parser::ParseOpenACCGangArg(SourceL...
[truncated]
|
DanielCChen
pushed a commit
to DanielCChen/llvm-project
that referenced
this pull request
Oct 16, 2024
The 'gang' clause is used to specify parallel execution of loops, thus has some complicated rules depending on the 'loop's associated compute construct. This patch implements all of those.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
The 'gang' clause is used to specify parallel execution of loops, thus has some complicated rules depending on the 'loop's associated compute construct. This patch implements all of those.