Per-dimension axis information for an SSA vector value of integer / index type. More...

Public Member Functions
	AxisInfo ()=default
unsigned	getRank () const
bool	isInitialized () const
bool	operator== (const AxisInfo &rhs) const
void	print (raw_ostream &os) const

Static Public Member Functions
static AxisInfo	getPessimistic (unsigned rank)
static AxisInfo	join (const AxisInfo &lhs, const AxisInfo &rhs)
	Conservative join (lattice meet).

Public Attributes
SmallVector< int64_t >	contiguity
SmallVector< int64_t >	constancy
SmallVector< int64_t >	divisibility
std::optional< int64_t >	knownConstant
std::optional< int64_t >	innerStride

Detailed Description

Per-dimension axis information for an SSA vector value of integer / index type.

The fields describe, for each dimension d, the pattern of the values along d (examples use a 1-D vector, so d is the only / innermost dim):

contiguity[d]: longest run that increases by exactly 1. [0, 1, 2, 3] -> 4; [0, 1, 0, 1] -> 2.
constancy[d]: longest run of equal values. [5, 5, 5, 5] -> 4; [5, 5, 6, 6] -> 2.
divisibility[d]: a power-of-two divisor of every element. [8, 16, 24] -> 8; [3, 6, 9] -> 1.
knownConstant: the value, if the whole vector is one constant. dense<7> -> 7; [0, 1, 2] -> nullopt.
innerStride: if set, consecutive values along the innermost dim differ by this constant step. 1 is the contiguous case ([0,1,2,3]), 0 the all-equal case ([5,5,5,5]); any other value is a strided progression ([0,4,8,12] -> 4) that contiguity/constancy can't represent (both read 1). For a multi-dim vector each inner-dim slice is its own progression; their bases may differ, with the shared inner alignment in divisibility[innerDim].

Only contiguity[innerDim] is consumed when stamping, but all dimensions are tracked because vector.transpose / vector.shape_cast permute or move per-dim info between axes, so an intermediate value's outer dims can become the inner dim of a later value.

Pessimistic / entry value: contiguity=1, constancy=1, divisibility=1, innerStride absent.

Definition at line 87 of file XeGPUContiguityAnalysis.cpp.

Constructor & Destructor Documentation

◆ AxisInfo()

mlir::xegpu::detail::axis_dataflow::AxisInfo::AxisInfo ( )

default

Referenced by getPessimistic(), join(), and operator==().

Member Function Documentation

◆ getPessimistic()

AxisInfo mlir::xegpu::detail::axis_dataflow::AxisInfo::getPessimistic ( unsigned rank )

inlinestatic

Definition at line 96 of file XeGPUContiguityAnalysis.cpp.

References AxisInfo(), constancy, contiguity, and divisibility.

Referenced by mlir::xegpu::detail::axis_dataflow::entryStateFor().

◆ getRank()

unsigned mlir::xegpu::detail::axis_dataflow::AxisInfo::getRank ( ) const

inline

Definition at line 104 of file XeGPUContiguityAnalysis.cpp.

References contiguity.

Referenced by isInitialized().

◆ isInitialized()

bool mlir::xegpu::detail::axis_dataflow::AxisInfo::isInitialized ( ) const

inline

Definition at line 105 of file XeGPUContiguityAnalysis.cpp.

References getRank().

◆ join()

AxisInfo mlir::xegpu::detail::axis_dataflow::AxisInfo::join	(	const AxisInfo &	lhs,
		const AxisInfo &	rhs )

inlinestatic

Conservative join (lattice meet).

When a value can arrive from several paths (e.g. a block argument, or arith.select), only what holds on every path is safe to assume. So we keep the weaker fact per field: min of each run length (a run is only guaranteed as long as the shortest incoming one), gcd of divisibility, and a value/stride only when both sides agree. This is what makes the contiguity we later stamp sound rather than "undecidable" — it is the largest run guaranteed on all paths.

Definition at line 120 of file XeGPUContiguityAnalysis.cpp.

References AxisInfo(), constancy, contiguity, divisibility, innerStride, knownConstant, lhs, and rhs.