Namespaces
namespace	AttributeTrait
namespace	detail

Classes
class	ACCDeclareEnterOpConversion
	Pattern to erase acc.declare_enter and its associated acc.declare_exit. More...
class	ACCOpEraseConversion
	Pattern to simply erase an ACC op (for ops with no results). More...
class	ACCOpReplaceWithVarConversion
	Pattern to replace an ACC op with its var operand. More...
class	ACCParMappingPolicy
	Policy class that defines how OpenACC parallelism levels map to target-specific parallel dimension attributes. More...
class	ACCRegionUnwrapConversion
	Pattern to unwrap a region from an ACC op and erase the wrapper. More...
struct	ConstructResource
struct	CurrentDeviceIdResource
class	DefaultACCToGPUMappingPolicy
	Default policy that provides the standard GPU mapping: gang(dim:1) -> BlockX (gridDim.x / blockIdx.x) gang(dim:2) -> BlockY (gridDim.y / blockIdx.y) gang(dim:3) -> BlockZ (gridDim.z / blockIdx.z) worker -> ThreadY (blockDim.y / threadIdx.y) vector -> ThreadX (blockDim.x / threadIdx.x) seq -> Sequential. More...
class	OpenACCSupport
struct	RuntimeCounters
class	SharedMemoryBudget
	Tracks aligned byte consumption against a configurable shared memory cap. More...
class	VariableInfoAttr
	Base attribute class for language-specific variable information carried through the OpenACC type interface helpers. More...

Typedefs
using	ACCToGPUMappingPolicy
	Type alias for the GPU-specific mapping policy.
using	TypeSizeAndAlignment = std::pair<llvm::TypeSize, llvm::TypeSize>
using	TypesForDevice = llvm::SmallSetVector<int64_t, 3>
	Holds information for which integers represent a device type in the runtime.

Enumerations
enum	OpenACCExecMapping { NONE = 0 , VECTOR = 1 , WORKER = 2 , GANG = 4 }
	Enumeration used to encode the execution mapping on a loop construct. More...

Functions
mlir::Value	getVar (mlir::Operation *accDataClauseOp)
	Used to obtain the var from a data clause operation.
mlir::TypedValue< mlir::acc::PointerLikeType >	getVarPtr (mlir::Operation *accDataClauseOp)
	Used to obtain the var from a data clause operation if it implements PointerLikeType.
mlir::Type	getVarType (mlir::Operation *accDataClauseOp)
	Used to obtains the varType from a data clause operation which records the type of variable.
mlir::Value	getAccVar (mlir::Operation *accDataClauseOp)
	Used to obtain the accVar from a data clause operation.
mlir::TypedValue< mlir::acc::PointerLikeType >	getAccPtr (mlir::Operation *accDataClauseOp)
	Used to obtain the accVar from a data clause operation if it implements PointerLikeType.
mlir::Value	getVarPtrPtr (mlir::Operation *accDataClauseOp)
	Used to obtain the varPtrPtr from a data clause operation.
mlir::SmallVector< mlir::Value >	getBounds (mlir::Operation *accDataClauseOp)
	Used to obtain bounds from an acc data clause operation.
mlir::SmallVector< mlir::Value >	getAsyncOperands (mlir::Operation *accDataClauseOp)
	Used to obtain async operands from an acc data clause operation.
mlir::ArrayAttr	getAsyncOperandsDeviceType (mlir::Operation *accDataClauseOp)
	Returns an array of acc:DeviceTypeAttr attributes attached to an acc data clause operation, that correspond to the device types associated with the async clauses with an async-value.
mlir::ArrayAttr	getAsyncOnly (mlir::Operation *accDataClauseOp)
	Returns an array of acc:DeviceTypeAttr attributes attached to an acc data clause operation, that correspond to the device types associated with the async clauses without an async-value.
std::optional< llvm::StringRef >	getVarName (mlir::Operation *accOp)
	Used to obtain the name from an acc operation.
std::optional< mlir::acc::DataClause >	getDataClause (mlir::Operation *accDataEntryOp)
	Used to obtain the dataClause from a data entry operation.
bool	getImplicitFlag (mlir::Operation *accDataEntryOp)
	Used to find out whether data operation is implicit.
mlir::ValueRange	getDataOperands (mlir::Operation *accOp)
	Used to get an immutable range iterating over the data operands.
mlir::MutableOperandRange	getMutableDataOperands (mlir::Operation *accOp)
	Used to get a mutable range iterating over the data operands.
mlir::SymbolRefAttr	getRecipe (mlir::Operation *accOp)
	Used to get the recipe attribute from a data clause operation.
bool	isPointerLikeType (mlir::Type type)
	Used to check whether the provided type implements the PointerLikeType interface.
bool	isMappableType (mlir::Type type)
	Used to check whether the provided type implements the MappableType interface.
static constexpr StringLiteral	getDeclareAttrName ()
	Used to obtain the attribute name for declare.
static constexpr StringLiteral	getDeclareActionAttrName ()
static constexpr StringLiteral	getRoutineInfoAttrName ()
static constexpr StringLiteral	getSpecializedRoutineAttrName ()
bool	isAccRoutine (mlir::Operation *op)
	Used to check whether the current operation is marked with acc routine.
bool	isSpecializedAccRoutine (mlir::Operation *op)
	Used to check whether this is a specialized accelerator version of acc routine function.
static constexpr StringLiteral	getFromDefaultClauseAttrName ()
static constexpr StringLiteral	getCollapseCountAttrName ()
	Name for an attribute attached to a loop indicating the number of loops collapsed to create that loop.
static constexpr StringLiteral	getVarNameAttrName ()
static constexpr StringLiteral	getCombinedConstructsAttrName ()
template<typename ComputeOpT>
static bool	isGangWorkerVectorAllOne (ComputeOpT op)
ParLevel	getGangParLevel (int64_t gangDimValue)
	Convert a gang dimension value (1, 2, or 3) to the corresponding ParLevel.
mlir::Operation *	getEnclosingComputeOp (mlir::Region &region)
	Used to obtain the enclosing compute construct operation that contains the provided region.
mlir::Value	getACCOperandForBlockArg (mlir::Value v)
	If v is not a block argument of an acc.compute_region body, returns nullptr.
mlir::Operation *	getACCDataClauseOpForBlockArg (mlir::Value v)
	If v is not a block argument of an acc.compute_region body, returns nullptr.
bool	isOnlyUsedByPrivateClauses (mlir::Value val, mlir::Region &region)
	Returns true if this value is only used by acc.private operations in the region.
bool	isOnlyUsedByReductionClauses (mlir::Value val, mlir::Region &region)
	Returns true if this value is only used by acc.reduction operations in the region.
std::optional< ClauseDefaultValue >	getDefaultAttr (mlir::Operation *op)
	Looks for an OpenACC default attribute on the current operation op or in a parent operation which encloses op.
mlir::acc::VariableTypeCategory	getTypeCategory (mlir::Value var)
	Get the type category of an OpenACC variable.
std::string	getVariableName (mlir::Value v)
	Attempts to extract the variable name from a value by walking through view-like operations until an acc.var_name attribute is found.
llvm::StringLiteral	getVarNamePlaceholder ()
	Returns a placeholder string for use as an acc.var_name attribute value when the actual variable name is not yet known at the point of IR construction.
std::string	getRecipeName (mlir::acc::RecipeKind kind, mlir::Type type)
	Get the recipe name for a given recipe kind and type.
mlir::Value	getBaseEntity (mlir::Value val)
bool	isValidSymbolUse (mlir::Operation user, mlir::SymbolRefAttr symbol, mlir::Operation *definingOpPtr=nullptr)
	Check if a symbol use is valid for use in an OpenACC region.
bool	isDeviceValue (mlir::Value val)
	Check if a value represents device data.
bool	isValidValueUse (mlir::Value val, mlir::Region &region)
	Check if a value use is valid in an OpenACC region.
llvm::SmallVector< mlir::Value >	getDominatingDataClauses (mlir::Operation *computeConstructOp, mlir::DominanceInfo &domInfo, mlir::PostDominanceInfo &postDomInfo)
	Collects all data clauses that dominate the compute construct.
remark::detail::InFlightRemark	emitRemark (mlir::Operation *op, const std::function< std::string()> &messageFn, llvm::StringRef category="openacc")
	Emit an OpenACC remark with lazy message generation.
remark::detail::InFlightRemark	emitRemark (mlir::Operation *op, const llvm::Twine &message, llvm::StringRef category="openacc")
	Emit an OpenACC remark for the given operation with the given message.
std::optional< DataLayout >	getDataLayout (Operation *op, bool allowDefault=true)
	Get the data layout for an operation.
ComputeRegionOp	buildComputeRegion (Location loc, ValueRange launchArgs, ValueRange inputArgs, llvm::StringRef origin, Region &regionToClone, RewriterBase &rewriter, IRMapping &mapping, ValueRange output={}, FlatSymbolRefAttr kernelFuncName={}, FlatSymbolRefAttr kernelModuleName={}, Value stream={}, ValueRange inputArgsToMap={})
	Build an acc.compute_region operation by cloning a source region.
void	insertParDim (llvm::SmallVector< GPUParallelDimAttr > &parDims, GPUParallelDimAttr parDim)
	Insert `parDim` into `parDims` while preserving dimension ordering.
void	removeParDim (llvm::SmallVector< GPUParallelDimAttr > &parDims, GPUParallelDimAttr parDim)
	Remove `parDim` from `parDims` if present.
GPUParallelDimsAttr	getParDimsAttr (Operation *op)
	Obtain the parallel dimensions carried by `op`, if any.
bool	hasParDimsAttr (Operation *op)
	Return whether `op` carries parallel dimensions.
bool	hasSeqParDims (Operation *op)
	Return whether `op` carries sequential parallel dimensions.
void	setParDimsAttr (Operation *op, GPUParallelDimsAttr attr)
	Set parallel dimensions on `op`.
void	updateParDimsAttr (Operation *op, GPUParallelDimsAttr attr)
	Update parallel dimensions on `op`.
void	copyParDimsAttr (Operation from, Operation to)
	Copy parallel dimensions from `from` to `to`.
bool	hasGPUBlockRedundantAttr (Operation *op)
	Return whether `op` is marked with the acc.gpu_block_redundant attribute, i.e.
void	setGPUBlockRedundantAttr (Operation *op)
	Mark `op` with the acc.gpu_block_redundant attribute.
GPUParallelDimsAttr	getGangDim1ParDimsAttr (MLIRContext *ctx, ACCToGPUMappingPolicy &policy)
	Create a gang dim 1 GPUParallelDimsAttr based on the mapping policy.
GPUParallelDimsAttr	getSeqParDimsAttr (MLIRContext *ctx, ACCToGPUMappingPolicy &policy)
	Create a sequential GPUParallelDimsAttr based on the mapping policy.
int64_t	sumExistingSharedMemoryBytes (Region &region)
	Sum aligned static_upper_bound_bytes for all acc.gpu_shared_memory in `region`.
PrivatizeOp	getPrivatizeOp (PrivateLocalOp privateLocal, ComputeRegionOp computeRegion)
	Resolve the acc.privatize operation associated with a private local.
MemRefType	getPrivateBaseMemRefType (Type baseTy, ModuleOp module)
	Returns the ranked MemRef type used to allocate privatized storage.
SmallVector< GPUParallelDimAttr >	collectPrivateLocalParDims (PrivateLocalOp privateLocal, ComputeRegionOp computeRegion)
	Collect parallel dimensions that govern privatization of `privateLocal`.
FailureOr< bool >	isPrivateLocalSharedMemoryCandidate (PrivateLocalOp privateLocal, ComputeRegionOp computeRegion, ModuleOp module, const ACCToGPUMappingPolicy &policy, OpenACCSupport *support=nullptr)
	True when `privateLocal` may be placed in shared memory.
std::optional< int64_t >	getPrivateLocalSharedMemoryUpperBoundBytes (PrivateLocalOp privateLocal, ComputeRegionOp computeRegion, ModuleOp module, const ACCToGPUMappingPolicy &policy, OpenACCSupport *support=nullptr)
	Upper-bound byte size for a shared-memory private_local candidate, or std::nullopt when not eligible or not statically computable.
std::optional< gpu::GPUModuleOp >	getOrCreateGPUModule (ModuleOp mod, bool create=true, llvm::StringRef name=kDefaultGPUModuleName)
	Get or create a GPU module in the given module.
Value	getGPUSize (gpu::Processor processor, gpu::LaunchOp launch, const llvm::DenseMap< gpu::Processor, Value > &dimensionOps)
	Return the launch dimension for `processor` from `launch`, or from `dimensionOps` when `launch` is null.
Value	getGPUThreadId (gpu::Processor processor, gpu::LaunchOp launch, const llvm::DenseMap< gpu::Processor, Value > &indexOps)
	Return the thread/block index for `processor` from `launch`, or from `indexOps` when `launch` is null.
std::pair< llvm::SmallVector< Value >, Block::iterator >	cloneACCRegionInto (Region src, Block dest, Block::iterator inlinePoint, IRMapping &mapping, ValueRange resultsToReplace)
	Clone an ACC region into a destination block at the given insertion point.
scf::ExecuteRegionOp	wrapMultiBlockRegionWithSCFExecuteRegion (Region &region, IRMapping &mapping, Location loc, RewriterBase &rewriter)
	Wrap a multi-block region in an scf.execute_region.
scf::ForOp	convertACCLoopToSCFFor (LoopOp loopOp, RewriterBase &rewriter, bool enableCollapse)
	Convert a structured acc.loop to scf.for.
scf::ParallelOp	convertACCLoopToSCFParallel (LoopOp loopOp, RewriterBase &rewriter)
	Convert acc.loop to scf.parallel.
scf::ExecuteRegionOp	convertUnstructuredACCLoopToSCFExecuteRegion (LoopOp loopOp, RewriterBase &rewriter)
	Convert an unstructured acc.loop to scf.execute_region.
void	setCollapseCountAttr (Operation *op, uint64_t count)
	Record on a collapsed loop how many original loops were folded into it.
uint64_t	getCollapseCount (Operation *op)
	Number of original loops collapsed into op, or 1 when op carries no collapse_count attribute.
SmallVector< GPUParallelDimAttr >	getReductionCombineParDims (ReductionCombineOp op)
	Returns the parallel dimensions that participate in `op's` combine step.
SmallVector< GPUParallelDimAttr >	getReductionCombineParDims (ReductionCombineRegionOp op)
	Returns the parallel dimensions that participate in `op's` combine step.
ReductionOperator	translateAtomicRMWKind (arith::AtomicRMWKind kind)
	Maps an arith atomic RMW kind to the corresponding acc reduction operator.
std::optional< arith::AtomicRMWKind >	translateACCReductionOperator (ReductionOperator redOp, Type type)
	Maps an acc reduction operator to the arith atomic RMW kind for `type`.
Value	createIdentityValue (OpBuilder &b, Location loc, Type type, arith::AtomicRMWKind kind, bool useOnlyFiniteValue=true)
	Creates the identity (neutral) value for a reduction of `type` and `kind`.
Value	generateReductionOp (OpBuilder &b, Location loc, Value lhs, Value rhs, arith::AtomicRMWKind kind)
	Combines two reduction partial values using the operator for `kind`.
llvm::SmallVector< mlir::acc::LoopOp >	uncollapseLoops (mlir::acc::LoopOp origLoop, unsigned tileCount, unsigned collapseCount, mlir::RewriterBase &rewriter)
	Uncollapse tile loops with multiple IVs and collapseCount < tileCount.
mlir::acc::LoopOp	tileACCLoops (llvm::SmallVector< mlir::acc::LoopOp > &tileLoops, const llvm::SmallVector< mlir::Value > &tileSizes, int32_t defaultTileSize, mlir::RewriterBase &rewriter)
	Tile ACC loops according to the given tile sizes.
std::optional< TypeSizeAndAlignment >	getTypeSizeAndAlignment (Type ty, ModuleOp module, const DataLayout &dl, OpenACCSupport *support=nullptr)
	Returns the size and ABI alignment in bytes.
std::optional< TypeSizeAndAlignment >	getTypeSizeAndAlignment (Type ty, ModuleOp module, OpenACCSupport *support=nullptr)
	Same as above, obtaining `dl` from `module` via getDataLayout.
void	populateACCSpecializeForDevicePatterns (RewritePatternSet &patterns, const TypesForDevice &theDeviceTypes)
	Populates all patterns for device specialization.
void	populateACCOrphanToHostPatterns (RewritePatternSet &patterns, OpenACCSupport &accSupport, bool enableLoopConversion=true)
	Populates patterns for converting orphan ACC operations to host.
void	populateACCHostFallbackPatterns (RewritePatternSet &patterns, OpenACCSupport &accSupport, bool enableLoopConversion=true)
	Populates all patterns for host fallback path (when if clause evaluates to false).
static SmallVector< GPUParallelDimAttr >::iterator	findParDim (SmallVector< GPUParallelDimAttr > &parDims, GPUParallelDimAttr parDim)
static bool	isThreadXPrivatize (PrivatizeOp privatize)
static FailureOr< std::optional< int64_t > >	getWorkerPrivateSharedMemoryNumCopies (PrivateLocalOp privateLocal, ComputeRegionOp computeRegion, bool isWorkerPrivate, OpenACCSupport *support)
static bool	isInsideACCSpecializedRoutine (Operation *op)
static Value	getGPUSizeFromLaunch (gpu::LaunchOp launch, gpu::Processor processor)
static Value	getGPUThreadIdFromLaunch (gpu::LaunchOp launch, gpu::Processor processor)
static bool	isFloatOrComplexType (Type ty)
static TypedAttr	getReductionIdentityValueAttr (arith::AtomicRMWKind kind, Type type, OpBuilder &builder, Location loc, bool useOnlyFiniteValue)
static std::optional< TypeSizeAndAlignment >	getTypeSizeAndAlignmentHelper (Type ty, ModuleOp module, const DataLayout &dl, OpenACCSupport *support)

Variables
constexpr llvm::StringLiteral	kDefaultGPUModuleName = "acc_gpu_module"
	Default GPU module name used by OpenACC.

Typedef Documentation

◆ ACCToGPUMappingPolicy

using mlir::acc::ACCToGPUMappingPolicy

Initial value:

ACCParMappingPolicy<mlir::acc::GPUParallelDimAttr>

mlir::acc::ACCParMappingPolicy

Policy class that defines how OpenACC parallelism levels map to target-specific parallel dimension at...

Definition OpenACCParMapping.h:64

Type alias for the GPU-specific mapping policy.

Definition at line 158 of file OpenACCParMapping.h.

◆ TypesForDevice

using mlir::acc::TypesForDevice = llvm::SmallSetVector<int64_t, 3>

Holds information for which integers represent a device type in the runtime.

Definition at line 33 of file Passes.h.

◆ TypeSizeAndAlignment

using mlir::acc::TypeSizeAndAlignment = std::pair<llvm::TypeSize, llvm::TypeSize>

Definition at line 29 of file OpenACCUtilsType.h.

Enumeration Type Documentation

◆ OpenACCExecMapping

enum mlir::acc::OpenACCExecMapping

Enumeration used to encode the execution mapping on a loop construct.

They refer directly to the OpenACC 3.3 standard: 2.9.2. gang 2.9.3. worker 2.9.4. vector

Value can be combined bitwise to reflect the mapping applied to the construct. e.g. acc.loop gang vector, the gang and vector could be combined and the final mapping value would be 5 (4 | 1).

Enumerator
NONE
VECTOR
WORKER
GANG

Definition at line 91 of file OpenACC.h.

Function Documentation

◆ buildComputeRegion()

ComputeRegionOp mlir::acc::buildComputeRegion	(	Location	loc,
		ValueRange	launchArgs,
		ValueRange	inputArgs,
		llvm::StringRef	origin,
		Region &	regionToClone,
		RewriterBase &	rewriter,
		IRMapping &	mapping,
		ValueRange	output = {},
		FlatSymbolRefAttr	kernelFuncName = {},
		FlatSymbolRefAttr	kernelModuleName = {},
		Value	stream = {},
		ValueRange	inputArgsToMap = {} )

Build an acc.compute_region operation by cloning a source region.

Creates a new acc.compute_region with the given launch arguments and origin string, then clones the operations from regionToClone into its body. Launch operands should be acc.par_width results (index); the region entry block gets matching index block arguments first, then arguments for each ins operand. Multi-block regions are wrapped with scf.execute_region.

The mapping is used and updated during cloning, allowing callers to track value correspondences. Optional output, kernelFuncName, kernelModuleName, and stream arguments are forwarded to the op.

When inputArgsToMap is non-empty, it is used as the key set for the clone mapping (instead of inputArgs). Use this when cloning a region that references one set of values (e.g. the source function's args) while the op's operands are another set (e.g. the current block's args). inputArgsToMap must have the same size as inputArgs when provided.

Definition at line 69 of file OpenACCUtilsCG.cpp.

References mlir::Block::addArgument(), mlir::OpBuilder::clone(), mlir::OpBuilder::createBlock(), mlir::RewriterBase::eraseOp(), mlir::Region::front(), mlir::Block::getArgument(), mlir::Region::getBlocks(), mlir::Builder::getIndexType(), mlir::Block::getOperations(), mlir::IRMapping::lookup(), mlir::IRMapping::map(), mlir::OpBuilder::setInsertionPointToEnd(), mlir::OpBuilder::setInsertionPointToStart(), and wrapMultiBlockRegionWithSCFExecuteRegion().

◆ cloneACCRegionInto()

std::pair< SmallVector< Value >, Block::iterator > mlir::acc::cloneACCRegionInto	(	Region *	src,
		Block *	dest,
		Block::iterator	inlinePoint,
		IRMapping &	mapping,
		ValueRange	resultsToReplace )

Clone an ACC region into a destination block at the given insertion point.

Requires a single-block source region. Maps block arguments and optional result replacement: values in resultsToReplace are replaced with the operands of the cloned region's acc.yield (1:1). Erases acc.yield/terminator and merges blocks. Returns (replacement values, insertion point after clone).

Definition at line 106 of file OpenACCUtilsLoop.cpp.

References mlir::Region::cloneInto(), mlir::Block::end(), mlir::Block::erase(), mlir::Block::getOperations(), mlir::Block::getParent(), mlir::Block::getTerminator(), mlir::Region::hasOneBlock(), mlir::replaceAllUsesInRegionWith(), replacement(), and mlir::Block::splitBlock().

◆ collectPrivateLocalParDims()

SmallVector< GPUParallelDimAttr > mlir::acc::collectPrivateLocalParDims	(	PrivateLocalOp	privateLocal,
		ComputeRegionOp	computeRegion )

Collect parallel dimensions that govern privatization of privateLocal.

Definition at line 277 of file OpenACCUtilsCG.cpp.

References getParDimsAttr(), getReductionCombineParDims(), mlir::Operation::getResult(), mlir::Value::getUsers(), and insertParDim().

Referenced by getPrivateLocalSharedMemoryUpperBoundBytes(), and isPrivateLocalSharedMemoryCandidate().

◆ convertACCLoopToSCFFor()

scf::ForOp mlir::acc::convertACCLoopToSCFFor	(	LoopOp	loopOp,
		RewriterBase &	rewriter,
		bool	enableCollapse )

Convert a structured acc.loop to scf.for.

The loop arguments are converted to index type. If enableCollapse is true, nested loops are collapsed into a single loop.

Parameters

loopOp	The acc.loop operation to convert (must not be unstructured)
rewriter	RewriterBase for creating operations
enableCollapse	Whether to collapse nested loops into one

Returns: The created scf.for operation or nullptr on creation error. An InFlightDiagnostic is emitted on creation error.

Definition at line 183 of file OpenACCUtilsLoop.cpp.

References mlir::coalesceLoops(), mlir::arith::ConstantIndexOp::create(), mlir::OpBuilder::getInsertionPoint(), mlir::IRMapping::map(), setCollapseCountAttr(), mlir::OpBuilder::setInsertionPoint(), mlir::OpBuilder::setInsertionPointToStart(), and mlir::Value::use_empty().

◆ convertACCLoopToSCFParallel()

scf::ParallelOp mlir::acc::convertACCLoopToSCFParallel	(	LoopOp	loopOp,
		RewriterBase &	rewriter )

Convert acc.loop to scf.parallel.

The loop induction variables are converted to index types.

Parameters

loopOp	The acc.loop operation to convert
rewriter	RewriterBase for creating and erasing operations

Returns: The created scf.parallel operation or nullptr on creation error. An InFlightDiagnostic is emitted on creation error.

Definition at line 263 of file OpenACCUtilsLoop.cpp.

References mlir::arith::ConstantIndexOp::create(), mlir::RewriterBase::eraseOp(), mlir::OpBuilder::getInsertionBlock(), mlir::OpBuilder::getInsertionPoint(), mlir::Block::getParentOp(), mlir::OpBuilder::setInsertionPointToStart(), mlir::Value::use_empty(), and wrapMultiBlockRegionWithSCFExecuteRegion().

◆ convertUnstructuredACCLoopToSCFExecuteRegion()

scf::ExecuteRegionOp mlir::acc::convertUnstructuredACCLoopToSCFExecuteRegion	(	LoopOp	loopOp,
		RewriterBase &	rewriter )

Convert an unstructured acc.loop to scf.execute_region.

Parameters

loopOp	The acc.loop operation to convert (must be unstructured)
rewriter	RewriterBase for creating and erasing operations

Returns: The created scf.execute_region operation or nullptr on creation error. An InFlightDiagnostic is emitted on creation error.

Definition at line 327 of file OpenACCUtilsLoop.cpp.

References mlir::OpBuilder::getInsertionBlock(), mlir::Block::getParentOp(), and wrapMultiBlockRegionWithSCFExecuteRegion().

◆ copyParDimsAttr()

void mlir::acc::copyParDimsAttr	(	Operation *	from,
		Operation *	to )

Copy parallel dimensions from from to to.

Definition at line 221 of file OpenACCUtilsCG.cpp.

References getParDimsAttr(), hasParDimsAttr(), and setParDimsAttr().

◆ createIdentityValue()

Value mlir::acc::createIdentityValue	(	OpBuilder &	b,
		Location	loc,
		Type	type,
		arith::AtomicRMWKind	kind,
		bool	useOnlyFiniteValue = true )

Creates the identity (neutral) value for a reduction of type and kind.

When useOnlyFiniteValue is true, floating-point identities avoid non-finite sentinel values where applicable.

Definition at line 186 of file OpenACCUtilsReduction.cpp.

References b, and getReductionIdentityValueAttr().

◆ emitRemark() [1/2]

remark::detail::InFlightRemark mlir::acc::emitRemark	(	mlir::Operation *	op,
		const llvm::Twine &	message,
		llvm::StringRef	category = "openacc" )

inline

Emit an OpenACC remark for the given operation with the given message.

Parameters

op	The operation to emit the remark for.
message	The remark message.
category	Optional category for the remark. Defaults to "openacc".

Returns: An in-flight remark object that can be used to append additional information to the remark.

Definition at line 150 of file OpenACCUtils.h.

References emitRemark().

◆ emitRemark() [2/2]

mlir::remark::detail::InFlightRemark mlir::acc::emitRemark	(	mlir::Operation *	op,
		const std::function< std::string()> &	messageFn,
		llvm::StringRef	category = "openacc" )

Emit an OpenACC remark with lazy message generation.

The messageFn is only invoked if remarks are enabled, allowing callers to avoid constructing expensive messages when remarks are disabled.

Parameters

op	The operation to emit the remark for.
messageFn	A callable that returns the remark message.
category	Optional category for the remark. Defaults to "openacc".

Returns: An in-flight remark object that can be used to append additional information to the remark.

Definition at line 371 of file OpenACCUtils.cpp.

References mlir::Attribute::getContext(), mlir::Operation::getLoc(), mlir::Operation::getParentOfType(), mlir::MLIRContext::getRemarkEngine(), and mlir::remark::RemarkOpts::name().

Referenced by mlir::acc::detail::OpenACCSupportTraits::Model< ImplT >::emitRemark(), emitRemark(), and mlir::acc::OpenACCSupport::emitRemark().

◆ findParDim()

SmallVector< GPUParallelDimAttr >::iterator mlir::acc::findParDim	(	SmallVector< GPUParallelDimAttr > &	parDims,
		GPUParallelDimAttr	parDim )

static

Definition at line 143 of file OpenACCUtilsCG.cpp.

References lhs, and rhs.

Referenced by insertParDim(), and removeParDim().

◆ generateReductionOp()

Value mlir::acc::generateReductionOp	(	OpBuilder &	b,
		Location	loc,
		Value	lhs,
		Value	rhs,
		arith::AtomicRMWKind	kind )

Combines two reduction partial values using the operator for kind.

Definition at line 203 of file OpenACCUtilsReduction.cpp.

References b, mlir::arith::getReductionOp(), lhs, and rhs.

◆ getACCDataClauseOpForBlockArg()

mlir::Operation * mlir::acc::getACCDataClauseOpForBlockArg ( mlir::Value v )

If v is not a block argument of an acc.compute_region body, returns nullptr.

Otherwise maps the block argument to its operand and returns the defining operation if it is one of ACC_DATA_ENTRY_OPS.

Definition at line 43 of file OpenACCUtils.cpp.

References getACCOperandForBlockArg(), and mlir::Value::getDefiningOp().

◆ getACCOperandForBlockArg()

mlir::Value mlir::acc::getACCOperandForBlockArg ( mlir::Value v )

If v is not a block argument of an acc.compute_region body, returns nullptr.

Otherwise maps the block argument to its operand and returns it.

Definition at line 28 of file OpenACCUtils.cpp.

References mlir::Block::getParentOp().

Referenced by getACCDataClauseOpForBlockArg().

◆ getAccPtr()

mlir::TypedValue< mlir::acc::PointerLikeType > mlir::acc::getAccPtr ( mlir::Operation * accDataClauseOp )

Used to obtain the accVar from a data clause operation if it implements PointerLikeType.

Definition at line 5249 of file OpenACC.cpp.

◆ getAccVar()

mlir::Value mlir::acc::getAccVar ( mlir::Operation * accDataClauseOp )

Used to obtain the accVar from a data clause operation.

When a data entry operation, it obtains its result accVar value. If a data exit operation, it obtains its operand accVar value. Returns empty value if not a data clause operation.

Definition at line 5261 of file OpenACC.cpp.

References ACC_DATA_ENTRY_OPS, and ACC_DATA_EXIT_OPS.

◆ getAsyncOnly()

mlir::ArrayAttr mlir::acc::getAsyncOnly ( mlir::Operation * accDataClauseOp )

Returns an array of acc:DeviceTypeAttr attributes attached to an acc data clause operation, that correspond to the device types associated with the async clauses without an async-value.

Definition at line 5316 of file OpenACC.cpp.

References ACC_DATA_ENTRY_OPS, and ACC_DATA_EXIT_OPS.

◆ getAsyncOperands()

mlir::SmallVector< mlir::Value > mlir::acc::getAsyncOperands ( mlir::Operation * accDataClauseOp )

Used to obtain async operands from an acc data clause operation.

Returns an empty vector if there are no such operands.

Definition at line 5294 of file OpenACC.cpp.

◆ getAsyncOperandsDeviceType()

mlir::ArrayAttr mlir::acc::getAsyncOperandsDeviceType ( mlir::Operation * accDataClauseOp )

Returns an array of acc:DeviceTypeAttr attributes attached to an acc data clause operation, that correspond to the device types associated with the async clauses with an async-value.

Definition at line 5308 of file OpenACC.cpp.

References ACC_DATA_ENTRY_OPS, and ACC_DATA_EXIT_OPS.

◆ getBaseEntity()

mlir::Value mlir::acc::getBaseEntity ( mlir::Value val )

Definition at line 182 of file OpenACCUtils.cpp.

References mlir::Value::getDefiningOp().

◆ getBounds()

mlir::SmallVector< mlir::Value > mlir::acc::getBounds ( mlir::Operation * accDataClauseOp )

Used to obtain bounds from an acc data clause operation.

Returns an empty vector if there are no bounds.

Definition at line 5279 of file OpenACC.cpp.

◆ getCollapseCount()

uint64_t mlir::acc::getCollapseCount ( Operation * op )

Number of original loops collapsed into op, or 1 when op carries no collapse_count attribute.

Definition at line 346 of file OpenACCUtilsLoop.cpp.

References mlir::Operation::getAttrOfType(), and getCollapseCountAttrName().

◆ getCollapseCountAttrName()

constexpr StringLiteral mlir::acc::getCollapseCountAttrName ( )

staticconstexpr

Name for an attribute attached to a loop indicating the number of loops collapsed to create that loop.

Definition at line 211 of file OpenACC.h.

Referenced by getCollapseCount(), and setCollapseCountAttr().

◆ getCombinedConstructsAttrName()

constexpr StringLiteral mlir::acc::getCombinedConstructsAttrName ( )

staticconstexpr

Definition at line 219 of file OpenACC.h.

◆ getDataClause()

std::optional< mlir::acc::DataClause > mlir::acc::getDataClause ( mlir::Operation * accDataEntryOp )

Used to obtain the dataClause from a data entry operation.

Returns empty optional if not a data entry operation.

Definition at line 5334 of file OpenACC.cpp.

Referenced by checkDeclareOperands().

◆ getDataLayout()

std::optional< DataLayout > mlir::acc::getDataLayout	(	Operation *	op,
		bool	allowDefault = true )

Get the data layout for an operation.

Attempts to get the data layout from the operation or its parent module. If allowDefault is true (default), a default data layout may be constructed when no explicit data layout spec is found.

Parameters

op	The operation to get the data layout for.
allowDefault	If true, allow returning a default data layout.

Returns: The data layout if available, std::nullopt otherwise.

Definition at line 34 of file OpenACCUtilsCG.cpp.

References mlir::Operation::getParentOfType(), and mlir::Operation::getParentOp().

Referenced by getTypeSizeAndAlignment().

◆ getDataOperands()

mlir::ValueRange mlir::acc::getDataOperands ( mlir::Operation * accOp )

Used to get an immutable range iterating over the data operands.

Definition at line 5352 of file OpenACC.cpp.

References ACC_COMPUTE_AND_DATA_CONSTRUCT_OPS.

◆ getDeclareActionAttrName()

constexpr StringLiteral mlir::acc::getDeclareActionAttrName ( )

staticconstexpr

Definition at line 181 of file OpenACC.h.

◆ getDeclareAttrName()

constexpr StringLiteral mlir::acc::getDeclareAttrName ( )

staticconstexpr

Used to obtain the attribute name for declare.

Definition at line 177 of file OpenACC.h.

Referenced by isDeviceValue(), and isValidSymbolUse().

◆ getDefaultAttr()

std::optional< mlir::acc::ClauseDefaultValue > mlir::acc::getDefaultAttr ( mlir::Operation * op )

Looks for an OpenACC default attribute on the current operation op or in a parent operation which encloses op.

This is useful because OpenACC specification notes that a visible default clause is the nearest default clause appearing on the compute construct or a lexically containing data construct.

Definition at line 75 of file OpenACCUtils.cpp.

References mlir::Operation::getParentOp().

◆ getDominatingDataClauses()

llvm::SmallVector< mlir::Value > mlir::acc::getDominatingDataClauses	(	mlir::Operation *	computeConstructOp,
		mlir::DominanceInfo &	domInfo,
		mlir::PostDominanceInfo &	postDomInfo )

Collects all data clauses that dominate the compute construct.

This includes data clauses from:

The compute construct itself
Enclosing data constructs
Applicable declare directives (those that dominate and post-dominate) This is used to determine if a variable is already covered by an existing data clause.
Parameters

computeConstructOp The compute construct operation

domInfo Dominance information

postDomInfo Post-dominance information

Returns
Vector of data clause values that dominate the compute construct

Definition at line 312 of file OpenACCUtils.cpp.

References mlir::DominanceInfo::dominates(), mlir::Operation::getParentOfType(), and mlir::Operation::getParentOp().

◆ getEnclosingComputeOp()

mlir::Operation * mlir::acc::getEnclosingComputeOp ( mlir::Region & region )

Used to obtain the enclosing compute construct operation that contains the provided region.

Returns nullptr if no compute construct operation is found. The returned operation is one of types defined by ACC_COMPUTE_CONSTRUCT_OPS.

Definition at line 23 of file OpenACCUtils.cpp.

References ACC_COMPUTE_CONSTRUCT_OPS, and mlir::Region::getParentOfType().

◆ getFromDefaultClauseAttrName()

constexpr StringLiteral mlir::acc::getFromDefaultClauseAttrName ( )

staticconstexpr

Definition at line 205 of file OpenACC.h.

◆ getGangDim1ParDimsAttr()

GPUParallelDimsAttr mlir::acc::getGangDim1ParDimsAttr	(	MLIRContext *	ctx,
		ACCToGPUMappingPolicy &	policy )

inline

Create a gang dim 1 GPUParallelDimsAttr based on the mapping policy.

Definition at line 110 of file OpenACCUtilsCG.h.

References mlir::acc::ACCParMappingPolicy< ParDimAttrT >::gangDim().

◆ getGangParLevel()

ParLevel mlir::acc::getGangParLevel ( int64_t gangDimValue )

inline

Convert a gang dimension value (1, 2, or 3) to the corresponding ParLevel.

Asserts if the value is not a valid gang dimension.

Definition at line 33 of file OpenACCParMapping.h.

◆ getGPUSize()

Value mlir::acc::getGPUSize	(	gpu::Processor	processor,
		gpu::LaunchOp	launch,
		const llvm::DenseMap< gpu::Processor, Value > &	dimensionOps )

Return the launch dimension for processor from launch, or from dimensionOps when launch is null.

Definition at line 68 of file OpenACCUtilsGPU.cpp.

References getGPUSizeFromLaunch().

◆ getGPUSizeFromLaunch()

Value mlir::acc::getGPUSizeFromLaunch	(	gpu::LaunchOp	launch,
		gpu::Processor	processor )

static

Definition at line 46 of file OpenACCUtilsGPU.cpp.

References mlir::gpu::KernelDim3::x, mlir::gpu::KernelDim3::y, and mlir::gpu::KernelDim3::z.

Referenced by getGPUSize().

◆ getGPUThreadId()

Value mlir::acc::getGPUThreadId	(	gpu::Processor	processor,
		gpu::LaunchOp	launch,
		const llvm::DenseMap< gpu::Processor, Value > &	indexOps )

Return the thread/block index for processor from launch, or from indexOps when launch is null.

Definition at line 98 of file OpenACCUtilsGPU.cpp.

References getGPUThreadIdFromLaunch().

◆ getGPUThreadIdFromLaunch()

Value mlir::acc::getGPUThreadIdFromLaunch	(	gpu::LaunchOp	launch,
		gpu::Processor	processor )

static

Definition at line 76 of file OpenACCUtilsGPU.cpp.

References mlir::gpu::KernelDim3::x, mlir::gpu::KernelDim3::y, and mlir::gpu::KernelDim3::z.

Referenced by getGPUThreadId().

◆ getImplicitFlag()

bool mlir::acc::getImplicitFlag ( mlir::Operation * accDataEntryOp )

Used to find out whether data operation is implicit.

Returns false if not a data operation or if it is a data operation without implicit flag.

Definition at line 5344 of file OpenACC.cpp.

References ACC_DATA_ENTRY_OPS.

◆ getMutableDataOperands()

mlir::MutableOperandRange mlir::acc::getMutableDataOperands ( mlir::Operation * accOp )

Used to get a mutable range iterating over the data operands.

Definition at line 5362 of file OpenACC.cpp.

References ACC_COMPUTE_AND_DATA_CONSTRUCT_OPS.

◆ getOrCreateGPUModule()

std::optional< gpu::GPUModuleOp > mlir::acc::getOrCreateGPUModule	(	ModuleOp	mod,
		bool	create = true,
		llvm::StringRef	name = kDefaultGPUModuleName )

Get or create a GPU module in the given module.

If a GPU module with the specified name already exists, it is returned. If create is true and no GPU module exists, one is created. If create is false and no GPU module exists, std::nullopt is returned.

Parameters

mod	The module to search or create the GPU module in.
create	If true (default), create the GPU module if it doesn't exist.
name	The name for the GPU module. If empty, uses kDefaultGPUModuleName.

Returns: The GPU module if found or created, std::nullopt otherwise.

Definition at line 20 of file OpenACCUtilsGPU.cpp.

References mlir::SymbolTable::insert(), kDefaultGPUModuleName, and mlir::SymbolTable::lookup().

Referenced by mlir::acc::detail::OpenACCSupportTraits::Model< ImplT >::getOrCreateGPUModule(), and mlir::acc::OpenACCSupport::getOrCreateGPUModule().

◆ getParDimsAttr()

GPUParallelDimsAttr mlir::acc::getParDimsAttr ( Operation * op )

Obtain the parallel dimensions carried by op, if any.

Definition at line 169 of file OpenACCUtilsCG.cpp.

References ACC_OP_WITH_PAR_DIMS_LIST, and mlir::Operation::getAttr().

Referenced by collectPrivateLocalParDims(), copyParDimsAttr(), getReductionCombineParDims(), getReductionCombineParDims(), hasParDimsAttr(), and hasSeqParDims().

◆ getPrivateBaseMemRefType()

MemRefType mlir::acc::getPrivateBaseMemRefType	(	Type	baseTy,
		ModuleOp	module )

Returns the ranked MemRef type used to allocate privatized storage.

baseTy is the baseTy parameter of acc.private_type (the privatized variable's type).

Definition at line 270 of file OpenACCUtilsCG.cpp.

Referenced by getPrivateLocalSharedMemoryUpperBoundBytes(), and isPrivateLocalSharedMemoryCandidate().

◆ getPrivateLocalSharedMemoryUpperBoundBytes()

std::optional< int64_t > mlir::acc::getPrivateLocalSharedMemoryUpperBoundBytes	(	PrivateLocalOp	privateLocal,
		ComputeRegionOp	computeRegion,
		ModuleOp	module,
		const ACCToGPUMappingPolicy &	policy,
		OpenACCSupport *	support = nullptr )

Upper-bound byte size for a shared-memory private_local candidate, or std::nullopt when not eligible or not statically computable.

Definition at line 411 of file OpenACCUtilsCG.cpp.

References collectPrivateLocalParDims(), getPrivateBaseMemRefType(), getTypeSizeAndAlignment(), getWorkerPrivateSharedMemoryNumCopies(), isPrivateLocalSharedMemoryCandidate(), and mlir::acc::ACCParMappingPolicy< ParDimAttrT >::isWorker().

◆ getPrivatizeOp()

PrivatizeOp mlir::acc::getPrivatizeOp	(	PrivateLocalOp	privateLocal,
		ComputeRegionOp	computeRegion )

Resolve the acc.privatize operation associated with a private local.

Definition at line 251 of file OpenACCUtilsCG.cpp.

References mlir::Value::getDefiningOp().

Referenced by isPrivateLocalSharedMemoryCandidate().

◆ getRecipe()

mlir::SymbolRefAttr mlir::acc::getRecipe ( mlir::Operation * accOp )

Used to get the recipe attribute from a data clause operation.

Definition at line 5371 of file OpenACC.cpp.

References ACC_DATA_ENTRY_OPS.

◆ getRecipeName()

std::string mlir::acc::getRecipeName	(	mlir::acc::RecipeKind	kind,
		mlir::Type	type )

Get the recipe name for a given recipe kind and type.

Returns an empty string if not possible to generate a recipe name.

Definition at line 144 of file OpenACCUtils.cpp.

References mlir::Type::print().

Referenced by mlir::acc::OpenACCSupport::getRecipeName().

◆ getReductionCombineParDims() [1/2]

SmallVector< GPUParallelDimAttr > mlir::acc::getReductionCombineParDims ( ReductionCombineOp op )

Returns the parallel dimensions that participate in op's combine step.

Used when lowering reductions to determine which GPU parallelism levels must be synchronized before combining partial results.

Definition at line 25 of file OpenACCUtilsReduction.cpp.

References getParDimsAttr().

Referenced by collectPrivateLocalParDims().

◆ getReductionCombineParDims() [2/2]

SmallVector< GPUParallelDimAttr > mlir::acc::getReductionCombineParDims ( ReductionCombineRegionOp op )

Returns the parallel dimensions that participate in op's combine step.

Prefers dimensions from an acc.reduction_accumulate user of the source variable; otherwise falls back to op's acc.par_dims attribute.

Definition at line 33 of file OpenACCUtilsReduction.cpp.

References getParDimsAttr(), and mlir::Operation::getUsers().

◆ getReductionIdentityValueAttr()

TypedAttr mlir::acc::getReductionIdentityValueAttr	(	arith::AtomicRMWKind	kind,
		Type	type,
		OpBuilder &	builder,
		Location	loc,
		bool	useOnlyFiniteValue )

static

Definition at line 141 of file OpenACCUtilsReduction.cpp.

References mlir::emitError(), mlir::arith::getIdentityValueAttr(), and mlir::Type::isIntOrIndexOrFloat().

Referenced by createIdentityValue().

◆ getRoutineInfoAttrName()

constexpr StringLiteral mlir::acc::getRoutineInfoAttrName ( )

staticconstexpr

Definition at line 185 of file OpenACC.h.

Referenced by isAccRoutine(), and isValidSymbolUse().

◆ getSeqParDimsAttr()

GPUParallelDimsAttr mlir::acc::getSeqParDimsAttr	(	MLIRContext *	ctx,
		ACCToGPUMappingPolicy &	policy )

inline

Create a sequential GPUParallelDimsAttr based on the mapping policy.

Definition at line 116 of file OpenACCUtilsCG.h.

References mlir::acc::ACCParMappingPolicy< ParDimAttrT >::seqDim().

◆ getSpecializedRoutineAttrName()

constexpr StringLiteral mlir::acc::getSpecializedRoutineAttrName ( )

staticconstexpr

Definition at line 189 of file OpenACC.h.

Referenced by isSpecializedAccRoutine().

◆ getTypeCategory()

mlir::acc::VariableTypeCategory mlir::acc::getTypeCategory ( mlir::Value var )

Get the type category of an OpenACC variable.

Definition at line 96 of file OpenACCUtils.cpp.

References mlir::Value::getType().

◆ getTypeSizeAndAlignment() [1/2]

std::optional< TypeSizeAndAlignment > mlir::acc::getTypeSizeAndAlignment	(	Type	ty,
		ModuleOp	module,
		const DataLayout &	dl,
		OpenACCSupport *	support = nullptr )

Returns the size and ABI alignment in bytes.

For aggregate structures and arrays, padding between members or elements is not taken into account. The result is a close estimate suitable for early OpenACC layout decisions, but not a complete ABI guarantee. For final size computations, use LLVM materialized types.

When support is provided, aggregate element types are sized by recursing through support so dialect-specific implementations can handle nested types.

Returns std::nullopt when the size is not statically computable or the type is not supported.

Definition at line 30 of file OpenACCUtilsType.cpp.

References mlir::Type::getContext(), mlir::DataLayout::getTypeABIAlignment(), mlir::DataLayout::getTypeSize(), getTypeSizeAndAlignmentHelper(), and mlir::Type::isIntOrIndexOrFloat().

Referenced by getPrivateLocalSharedMemoryUpperBoundBytes(), mlir::acc::detail::OpenACCSupportTraits::Model< ImplT >::getTypeSizeAndAlignment(), getTypeSizeAndAlignment(), mlir::acc::OpenACCSupport::getTypeSizeAndAlignment(), and getTypeSizeAndAlignmentHelper().

◆ getTypeSizeAndAlignment() [2/2]

std::optional< TypeSizeAndAlignment > mlir::acc::getTypeSizeAndAlignment	(	Type	ty,
		ModuleOp	module,
		OpenACCSupport *	support = nullptr )

Same as above, obtaining dl from module via getDataLayout.

Definition at line 84 of file OpenACCUtilsType.cpp.

References getDataLayout(), and getTypeSizeAndAlignment().

◆ getTypeSizeAndAlignmentHelper()

std::optional< TypeSizeAndAlignment > mlir::acc::getTypeSizeAndAlignmentHelper	(	Type	ty,
		ModuleOp	module,
		const DataLayout &	dl,
		OpenACCSupport *	support )

static

Definition at line 22 of file OpenACCUtilsType.cpp.

References getTypeSizeAndAlignment(), and mlir::acc::OpenACCSupport::getTypeSizeAndAlignment().

Referenced by getTypeSizeAndAlignment().

◆ getVar()

mlir::Value mlir::acc::getVar ( mlir::Operation * accDataClauseOp )

Used to obtain the var from a data clause operation.

Returns empty value if not a data clause operation or is a data exit operation with no var.

Definition at line 5230 of file OpenACC.cpp.

References ACC_DATA_ENTRY_OPS.

Referenced by checkDeclareOperands().

◆ getVariableName()

std::string mlir::acc::getVariableName ( mlir::Value v )

Attempts to extract the variable name from a value by walking through view-like operations until an acc.var_name attribute is found.

Returns empty string if no name is found.

Definition at line 113 of file OpenACCUtils.cpp.

References getConstantIntValue(), mlir::Value::getDefiningOp(), getVarName(), and getVarNameAttrName().

Referenced by mlir::acc::OpenACCSupport::getVariableName().

◆ getVarName()

std::optional< llvm::StringRef > mlir::acc::getVarName ( mlir::Operation * accOp )

Used to obtain the name from an acc operation.

Definition at line 5323 of file OpenACC.cpp.

Referenced by getVariableName().

◆ getVarNameAttrName()

constexpr StringLiteral mlir::acc::getVarNameAttrName ( )

staticconstexpr

Definition at line 215 of file OpenACC.h.

Referenced by getVariableName().

◆ getVarNamePlaceholder()

llvm::StringLiteral mlir::acc::getVarNamePlaceholder ( )

Returns a placeholder string for use as an acc.var_name attribute value when the actual variable name is not yet known at the point of IR construction.

The placeholder is meant to be replaced with the real name at a later lowering stage. For example, recipe init regions may attach this to ops at recipe-generation time, and ACCRecipeMaterialization will subsequently replace the placeholder with the actual variable name on all marked ops after inlining the recipe into the compute construct.

Definition at line 109 of file OpenACCUtils.cpp.

◆ getVarPtr()

mlir::TypedValue< mlir::acc::PointerLikeType > mlir::acc::getVarPtr ( mlir::Operation * accDataClauseOp )

Used to obtain the var from a data clause operation if it implements PointerLikeType.

Definition at line 5216 of file OpenACC.cpp.

◆ getVarPtrPtr()

mlir::Value mlir::acc::getVarPtrPtr ( mlir::Operation * accDataClauseOp )

Used to obtain the varPtrPtr from a data clause operation.

Returns empty value if not a data clause operation.

Definition at line 5269 of file OpenACC.cpp.

References ACC_DATA_ENTRY_OPS.

◆ getVarType()

mlir::Type mlir::acc::getVarType ( mlir::Operation * accDataClauseOp )

Used to obtains the varType from a data clause operation which records the type of variable.

When var is PointerLikeType, this returns the type of the pointer target.

Definition at line 5238 of file OpenACC.cpp.

References ACC_DATA_ENTRY_OPS.

◆ getWorkerPrivateSharedMemoryNumCopies()

FailureOr< std::optional< int64_t > > mlir::acc::getWorkerPrivateSharedMemoryNumCopies	(	PrivateLocalOp	privateLocal,
		ComputeRegionOp	computeRegion,
		bool	isWorkerPrivate,
		OpenACCSupport *	support )

static

Definition at line 321 of file OpenACCUtilsCG.cpp.

References mlir::ValueBoundsConstraintSet::computeConstantBound(), mlir::acc::OpenACCSupport::emitNYI(), and mlir::presburger::EQ.

Referenced by getPrivateLocalSharedMemoryUpperBoundBytes(), and isPrivateLocalSharedMemoryCandidate().

◆ hasGPUBlockRedundantAttr()

bool mlir::acc::hasGPUBlockRedundantAttr ( Operation * op )

Return whether op is marked with the acc.gpu_block_redundant attribute, i.e.

it executes redundantly across all thread blocks. Such an op must not be assigned block/grid-level work-sharing; only thread-level parallelism may apply, and its enclosing block dimensions are treated as active (not predicated).

Definition at line 212 of file OpenACCUtilsCG.cpp.

References mlir::Operation::hasAttrOfType().

◆ hasParDimsAttr()

bool mlir::acc::hasParDimsAttr ( Operation * op )

Return whether op carries parallel dimensions.

Definition at line 183 of file OpenACCUtilsCG.cpp.

References getParDimsAttr().

Referenced by copyParDimsAttr(), setParDimsAttr(), and updateParDimsAttr().

◆ hasSeqParDims()

bool mlir::acc::hasSeqParDims ( Operation * op )

Return whether op carries sequential parallel dimensions.

Definition at line 185 of file OpenACCUtilsCG.cpp.

References getParDimsAttr().

◆ insertParDim()

void mlir::acc::insertParDim	(	llvm::SmallVector< GPUParallelDimAttr > &	parDims,
		GPUParallelDimAttr	parDim )

Insert parDim into parDims while preserving dimension ordering.

If the dimension is already present, this is a no-op.

Definition at line 152 of file OpenACCUtilsCG.cpp.

References findParDim().

Referenced by collectPrivateLocalParDims().

◆ isAccRoutine()

bool mlir::acc::isAccRoutine ( mlir::Operation * op )

inline

Used to check whether the current operation is marked with acc routine.

The operation passed in should be a function.

Definition at line 195 of file OpenACC.h.

References getRoutineInfoAttrName(), and mlir::Operation::hasAttr().

◆ isDeviceValue()

bool mlir::acc::isDeviceValue ( mlir::Value val )

Check if a value represents device data.

This checks if the value represents device data via the MappableType, PointerLikeType, and GlobalVariableOpInterface interfaces, and whether the defining operation carries acc.declare with the deviceptr clause.

Parameters

val	The value to check

Returns: true if the value is device data, false otherwise

Definition at line 246 of file OpenACCUtils.cpp.

References mlir::Operation::getAttrOfType(), getDeclareAttrName(), mlir::Value::getDefiningOp(), mlir::Value::getType(), isDeviceValue(), and mlir::SymbolTable::lookupNearestSymbolFrom().

Referenced by isDeviceValue(), and isValidValueUse().

◆ isFloatOrComplexType()

bool mlir::acc::isFloatOrComplexType ( Type ty )

static

Definition at line 20 of file OpenACCUtilsReduction.cpp.

Referenced by translateACCReductionOperator().

◆ isGangWorkerVectorAllOne()

template<typename ComputeOpT>

bool mlir::acc::isGangWorkerVectorAllOne ( ComputeOpT op )

static

Definition at line 242 of file OpenACC.h.

References mlir::isConstantIntValue().

◆ isInsideACCSpecializedRoutine()

bool mlir::acc::isInsideACCSpecializedRoutine ( Operation * op )

static

Definition at line 352 of file OpenACCUtilsCG.cpp.

References mlir::Operation::getParentOfType(), and isSpecializedAccRoutine().

Referenced by isPrivateLocalSharedMemoryCandidate().

◆ isMappableType()

bool mlir::acc::isMappableType ( mlir::Type type )

inline

Used to check whether the provided type implements the MappableType interface.

Definition at line 172 of file OpenACC.h.

Referenced by checkRecipe().

◆ isOnlyUsedByPrivateClauses()

bool mlir::acc::isOnlyUsedByPrivateClauses	(	mlir::Value	val,
		mlir::Region &	region )

Returns true if this value is only used by acc.private operations in the region.

Definition at line 64 of file OpenACCUtils.cpp.

References isOnlyUsedByOpClauses().

Referenced by isValidValueUse().

◆ isOnlyUsedByReductionClauses()

bool mlir::acc::isOnlyUsedByReductionClauses	(	mlir::Value	val,
		mlir::Region &	region )

Returns true if this value is only used by acc.reduction operations in the region.

Definition at line 69 of file OpenACCUtils.cpp.

References isOnlyUsedByOpClauses().

◆ isPointerLikeType()

bool mlir::acc::isPointerLikeType ( mlir::Type type )

inline

Used to check whether the provided type implements the PointerLikeType interface.

Definition at line 166 of file OpenACC.h.

◆ isPrivateLocalSharedMemoryCandidate()

FailureOr< bool > mlir::acc::isPrivateLocalSharedMemoryCandidate	(	PrivateLocalOp	privateLocal,
		ComputeRegionOp	computeRegion,
		ModuleOp	module,
		const ACCToGPUMappingPolicy &	policy,
		OpenACCSupport *	support = nullptr )

True when privateLocal may be placed in shared memory.

Definition at line 357 of file OpenACCUtilsCG.cpp.

References collectPrivateLocalParDims(), getPrivateBaseMemRefType(), getPrivatizeOp(), getWorkerPrivateSharedMemoryNumCopies(), mlir::acc::ACCParMappingPolicy< ParDimAttrT >::isGang(), isInsideACCSpecializedRoutine(), isThreadXPrivatize(), mlir::acc::ACCParMappingPolicy< ParDimAttrT >::isVector(), and mlir::acc::ACCParMappingPolicy< ParDimAttrT >::isWorker().

Referenced by getPrivateLocalSharedMemoryUpperBoundBytes().

◆ isSpecializedAccRoutine()

bool mlir::acc::isSpecializedAccRoutine ( mlir::Operation * op )

inline

Used to check whether this is a specialized accelerator version of acc routine function.

Definition at line 201 of file OpenACC.h.

References getSpecializedRoutineAttrName(), and mlir::Operation::hasAttr().

Referenced by isInsideACCSpecializedRoutine().

◆ isThreadXPrivatize()

bool mlir::acc::isThreadXPrivatize ( PrivatizeOp privatize )

static

Definition at line 263 of file OpenACCUtilsCG.cpp.

Referenced by isPrivateLocalSharedMemoryCandidate().

◆ isValidSymbolUse()

bool mlir::acc::isValidSymbolUse	(	mlir::Operation *	user,
		mlir::SymbolRefAttr	symbol,
		mlir::Operation **	definingOpPtr = nullptr )

Check if a symbol use is valid for use in an OpenACC region.

This includes looking for various attributes such as acc.routine_info and acc.declare attributes.

Parameters

user	The operation using the symbol
symbol	The symbol reference being used
definingOpPtr	Optional output parameter to receive the defining op

Returns: true if the symbol use is valid, false otherwise

Definition at line 192 of file OpenACCUtils.cpp.

References getDeclareAttrName(), getRoutineInfoAttrName(), mlir::Operation::hasAttr(), mlir::SymbolTable::lookupNearestSymbolFrom(), and mlir::SymbolTable::Private.

Referenced by mlir::acc::detail::OpenACCSupportTraits::Model< ImplT >::isValidSymbolUse(), and mlir::acc::OpenACCSupport::isValidSymbolUse().

◆ isValidValueUse()

bool mlir::acc::isValidValueUse	(	mlir::Value	val,
		mlir::Region &	region )

Check if a value use is valid in an OpenACC region.

This is true if:

The value is produced by an ACC data entry operation
The value is device data
The value is only used by private clauses in the region
Parameters

val The value to check

region The OpenACC region

Returns
true if the value use is valid, false otherwise

Definition at line 289 of file OpenACCUtils.cpp.

References mlir::Value::getDefiningOp(), mlir::Value::getType(), isDeviceValue(), mlir::Type::isIntOrIndexOrFloat(), and isOnlyUsedByPrivateClauses().

Referenced by mlir::acc::detail::OpenACCSupportTraits::Model< ImplT >::isValidValueUse(), and mlir::acc::OpenACCSupport::isValidValueUse().

◆ populateACCHostFallbackPatterns()

void mlir::acc::populateACCHostFallbackPatterns	(	RewritePatternSet &	patterns,
		OpenACCSupport &	accSupport,
		bool	enableLoopConversion = true )

Populates all patterns for host fallback path (when if clause evaluates to false).

In this mode, ALL ACC operations should be converted or removed.

Parameters

enableLoopConversion Whether to convert orphan acc.loop operations.

Definition at line 404 of file ACCSpecializeForHost.cpp.

References mlir::RewritePatternSet::getContext(), and mlir::RewritePatternSet::insert().

◆ populateACCOrphanToHostPatterns()

void mlir::acc::populateACCOrphanToHostPatterns	(	RewritePatternSet &	patterns,
		OpenACCSupport &	accSupport,
		bool	enableLoopConversion = true )

Populates patterns for converting orphan ACC operations to host.

All patterns check that the operation is NOT inside or associated with a compute region before converting.

Parameters

enableLoopConversion Whether to convert orphan acc.loop operations.

Definition at line 367 of file ACCSpecializeForHost.cpp.

References mlir::RewritePatternSet::getContext(), and mlir::RewritePatternSet::insert().

◆ populateACCSpecializeForDevicePatterns()

void mlir::acc::populateACCSpecializeForDevicePatterns	(	RewritePatternSet &	patterns,
		const TypesForDevice &	theDeviceTypes )

Populates all patterns for device specialization.

In specialized device code (such as specialized acc routine), many ACC operations do not make sense because they are host-side constructs. This function adds patterns to remove or transform them. theDeviceTypes hold the device types for which acc_on_device should return true.

Definition at line 161 of file ACCSpecializeForDevice.cpp.

References mlir::RewritePatternSet::getContext(), and mlir::RewritePatternSet::insert().

◆ removeParDim()

void mlir::acc::removeParDim	(	llvm::SmallVector< GPUParallelDimAttr > &	parDims,
		GPUParallelDimAttr	parDim )

Remove parDim from parDims if present.

Definition at line 159 of file OpenACCUtilsCG.cpp.

References findParDim().

◆ setCollapseCountAttr()

void mlir::acc::setCollapseCountAttr	(	Operation *	op,
		uint64_t	count )

Record on a collapsed loop how many original loops were folded into it.

Definition at line 341 of file OpenACCUtilsLoop.cpp.

References getCollapseCountAttrName(), mlir::Operation::getContext(), and mlir::Operation::setAttr().

Referenced by convertACCLoopToSCFFor().

◆ setGPUBlockRedundantAttr()

void mlir::acc::setGPUBlockRedundantAttr ( Operation * op )

Mark op with the acc.gpu_block_redundant attribute.

Definition at line 216 of file OpenACCUtilsCG.cpp.

References mlir::Operation::getContext(), and mlir::Operation::setAttr().

◆ setParDimsAttr()

void mlir::acc::setParDimsAttr	(	Operation *	op,
		GPUParallelDimsAttr	attr )

Set parallel dimensions on op.

Definition at line 191 of file OpenACCUtilsCG.cpp.

References ACC_OP_WITH_PAR_DIMS_LIST, hasParDimsAttr(), and mlir::Operation::setAttr().

Referenced by copyParDimsAttr().

◆ sumExistingSharedMemoryBytes()

int64_t mlir::acc::sumExistingSharedMemoryBytes ( Region & region )

Sum aligned static_upper_bound_bytes for all acc.gpu_shared_memory in region.

Definition at line 242 of file OpenACCUtilsCG.cpp.

References mlir::acc::SharedMemoryBudget::alignOffset(), and mlir::Region::walk().

◆ tileACCLoops()

mlir::acc::LoopOp mlir::acc::tileACCLoops	(	llvm::SmallVector< mlir::acc::LoopOp > &	tileLoops,
		const llvm::SmallVector< mlir::Value > &	tileSizes,
		int32_t	defaultTileSize,
		mlir::RewriterBase &	rewriter )

Tile ACC loops according to the given tile sizes.

Tiling a 2-level nested loop will create two 'tile' loops containing two 'element' loops. The transformation looks like:

Before Tiling:

#pragma acc loop tile(tile_size1, tile_size2)
 for (i = lb1; i < ub1; i += step1) { // original loop
   for (j = lb2; j < ub2; j += step2) {
     a[i,j] = i + j;
   }
 }

After Tiling:

for (i = lb1; i < ub1; i += (step1 * tile_size1)) { // tile loop 1
  for (j = lb2; j < ub2; j += (step2 * tile_size2)) { // tile loop 2
    for (ii = i; ii < min(ub1, (step1 * tile_size1) + i); ii += step1) {
    // element loop 1
      for (jj = j; jj < min(ub2, (step2 * tile_size2) + j); jj += step2)
      { // element loop 2
        a[ii,jj] = i + j;
      }
    }
  }
}

Unknown tile sizes (represented as -1 in acc dialect for tile(*)) are resolved to the provided default tile size.

Parameters

tileLoops	The loops to tile (outermost first).
tileSizes	The tile sizes for each dimension. Values of -1 are treated as unknown and resolved to defaultTileSize.
defaultTileSize	The default tile size to use for unknown (*) tiles.
rewriter	The rewriter to use for modifications.

Returns: The outermost loop after tiling.

Definition at line 163 of file OpenACCUtilsTiling.cpp.

References createInnerLoop(), mlir::RewriterBase::finalizeOpModification(), mlir::Builder::getDenseBoolArrayAttr(), mlir::Builder::getIntegerAttr(), mlir::getType(), moveOpsAndReplaceIVs(), removeWorkerVectorFromLoop(), resolveAndCastTileSize(), mlir::OpBuilder::setInsertionPoint(), and mlir::RewriterBase::startOpModification().

◆ translateACCReductionOperator()

std::optional< arith::AtomicRMWKind > mlir::acc::translateACCReductionOperator	(	ReductionOperator	redOp,
		Type	type )

Maps an acc reduction operator to the arith atomic RMW kind for type.

Returns std::nullopt when redOp is not supported for type.

Definition at line 75 of file OpenACCUtilsReduction.cpp.

References mlir::Type::isFloat(), isFloatOrComplexType(), mlir::Type::isInteger(), and mlir::Type::isUnsignedInteger().

◆ translateAtomicRMWKind()

ReductionOperator mlir::acc::translateAtomicRMWKind ( arith::AtomicRMWKind kind )

Maps an arith atomic RMW kind to the corresponding acc reduction operator.

Definition at line 44 of file OpenACCUtilsReduction.cpp.

◆ uncollapseLoops()

llvm::SmallVector< mlir::acc::LoopOp > mlir::acc::uncollapseLoops	(	mlir::acc::LoopOp	origLoop,
		unsigned	tileCount,
		unsigned	collapseCount,
		mlir::RewriterBase &	rewriter )

Uncollapse tile loops with multiple IVs and collapseCount < tileCount.

This is used to prepare loops for tiling when the collapse count is less than the tile count.

Parameters

origLoop	The original loop operation to uncollapse.
tileCount	The number of tile dimensions.
collapseCount	The collapse count from the original loop.
rewriter	The rewriter to use for modifications.

Returns: A vector of uncollapsed loop operations.

Definition at line 280 of file OpenACCUtilsTiling.cpp.

References createACCLoopFromOriginal(), mlir::OpBuilder::createBlock(), createInnerLoop(), mlir::Builder::getDenseBoolArrayAttr(), moveOpsAndReplaceIVs(), mlir::OpBuilder::setInsertionPoint(), and mlir::OpBuilder::setInsertionPointToEnd().

◆ updateParDimsAttr()

void mlir::acc::updateParDimsAttr	(	Operation *	op,
		GPUParallelDimsAttr	attr )

Update parallel dimensions on op.

Definition at line 200 of file OpenACCUtilsCG.cpp.

References ACC_OP_WITH_PAR_DIMS_LIST, hasParDimsAttr(), and mlir::Operation::setAttr().

◆ wrapMultiBlockRegionWithSCFExecuteRegion()

scf::ExecuteRegionOp mlir::acc::wrapMultiBlockRegionWithSCFExecuteRegion	(	Region &	region,
		IRMapping &	mapping,
		Location	loc,
		RewriterBase &	rewriter )

Wrap a multi-block region in an scf.execute_region.

Wrap a multi-block region with scf.execute_region.

Clones the given region into a new scf.execute_region. Terminators with no successors (i.e., region exit points) are replaced with scf.yield. Use this to convert unstructured control flow (e.g. multiple blocks with branches) into a single SCF region.

Parameters

region	The region to wrap (cloned into the execute_region; not modified).
mapping	IR mapping for the clone; updated with block and value mappings.
loc	Location for the created execute_region op.
rewriter	RewriterBase for creating and erasing operations.

Returns: The created scf.execute_region operation.

Definition at line 150 of file OpenACCUtilsLoop.cpp.

References mlir::OpBuilder::cloneRegionBefore(), mlir::Region::end(), mlir::RewriterBase::eraseOp(), mlir::Region::getBlocks(), mlir::Operation::getLoc(), mlir::Operation::getNumSuccessors(), mlir::Operation::getOperands(), mlir::IRMapping::lookup(), mlir::OpBuilder::setInsertionPoint(), and TypeRange.

Referenced by buildComputeRegion(), convertACCLoopToSCFParallel(), and convertUnstructuredACCLoopToSCFExecuteRegion().

Variable Documentation

◆ kDefaultGPUModuleName

llvm::StringLiteral mlir::acc::kDefaultGPUModuleName = "acc_gpu_module"

constexpr

Default GPU module name used by OpenACC.

Definition at line 27 of file OpenACCUtilsGPU.h.

Referenced by getOrCreateGPUModule().

computeConstructOp	The compute construct operation
domInfo	Dominance information
postDomInfo	Post-dominance information

Namespaces

Classes

Typedefs

Enumerations

Functions

Variables

Typedef Documentation

◆ ACCToGPUMappingPolicy

◆ TypesForDevice

◆ TypeSizeAndAlignment

Enumeration Type Documentation

◆ OpenACCExecMapping

Function Documentation

◆ buildComputeRegion()

◆ cloneACCRegionInto()

◆ collectPrivateLocalParDims()

◆ convertACCLoopToSCFFor()

◆ convertACCLoopToSCFParallel()

◆ convertUnstructuredACCLoopToSCFExecuteRegion()

◆ copyParDimsAttr()

◆ createIdentityValue()

◆ emitRemark() [1/2]

◆ emitRemark() [2/2]

◆ findParDim()

◆ generateReductionOp()

◆ getACCDataClauseOpForBlockArg()

◆ getACCOperandForBlockArg()

◆ getAccPtr()

◆ getAccVar()

◆ getAsyncOnly()

◆ getAsyncOperands()

◆ getAsyncOperandsDeviceType()

◆ getBaseEntity()

◆ getBounds()

◆ getCollapseCount()

◆ getCollapseCountAttrName()

◆ getCombinedConstructsAttrName()

◆ getDataClause()

◆ getDataLayout()

◆ getDataOperands()

◆ getDeclareActionAttrName()

◆ getDeclareAttrName()

◆ getDefaultAttr()

◆ getDominatingDataClauses()

◆ getEnclosingComputeOp()

◆ getFromDefaultClauseAttrName()

◆ getGangDim1ParDimsAttr()

◆ getGangParLevel()

◆ getGPUSize()

◆ getGPUSizeFromLaunch()

◆ getGPUThreadId()

◆ getGPUThreadIdFromLaunch()

◆ getImplicitFlag()

◆ getMutableDataOperands()

◆ getOrCreateGPUModule()

◆ getParDimsAttr()

◆ getPrivateBaseMemRefType()

◆ getPrivateLocalSharedMemoryUpperBoundBytes()

◆ getPrivatizeOp()

◆ getRecipe()

◆ getRecipeName()

◆ getReductionCombineParDims() [1/2]

◆ getReductionCombineParDims() [2/2]

◆ getReductionIdentityValueAttr()

◆ getRoutineInfoAttrName()

◆ getSeqParDimsAttr()

◆ getSpecializedRoutineAttrName()

◆ getTypeCategory()

◆ getTypeSizeAndAlignment() [1/2]

◆ getTypeSizeAndAlignment() [2/2]

◆ getTypeSizeAndAlignmentHelper()

◆ getVar()

◆ getVariableName()

◆ getVarName()

◆ getVarNameAttrName()

◆ getVarNamePlaceholder()

◆ getVarPtr()

◆ getVarPtrPtr()

◆ getVarType()

◆ getWorkerPrivateSharedMemoryNumCopies()