This class allows control over how the GreedyPatternRewriteDriver works. More...

#include "mlir/Transforms/GreedyPatternRewriteDriver.h"

Public Member Functions
bool	getUseTopDownTraversal () const
	This specifies the order of initial traversal that populates the rewriters worklist.
GreedyRewriteConfig &	setUseTopDownTraversal (bool use=true)
GreedySimplifyRegionLevel	getRegionSimplificationLevel () const
	Perform control flow optimizations to the region tree after applying all patterns.
GreedyRewriteConfig &	setRegionSimplificationLevel (GreedySimplifyRegionLevel level)
int64_t	getMaxIterations () const
	This specifies the maximum number of times the rewriter will iterate between applying patterns and simplifying regions.
GreedyRewriteConfig &	setMaxIterations (int64_t iterations)
int64_t	getMaxNumRewrites () const
	This specifies the maximum number of rewrites within an iteration.
GreedyRewriteConfig &	setMaxNumRewrites (int64_t limit)
Region *	getScope () const
	Only ops within the scope are added to the worklist.
GreedyRewriteConfig &	setScope (Region *scope)
GreedyRewriteStrictness	getStrictness () const
	Strict mode can restrict the ops that are added to the worklist during the rewrite.
GreedyRewriteConfig &	setStrictness (GreedyRewriteStrictness mode)
RewriterBase::Listener *	getListener () const
	An optional listener that should be notified about IR modifications.
GreedyRewriteConfig &	setListener (RewriterBase::Listener *listener)
bool	isFoldingEnabled () const
	Whether this should fold while greedily rewriting.
GreedyRewriteConfig &	enableFolding (bool enable=true)
bool	isConstantCSEEnabled () const
	If set to "true", constants are CSE'd (even across multiple regions that are in a parent-ancestor relationship).
GreedyRewriteConfig &	enableConstantCSE (bool enable=true)
bool	isCSEBetweenIterationsEnabled () const
	If set to "true", full common-subexpression elimination is run on the scoped region between each pattern-application iteration.
GreedyRewriteConfig &	enableCSEBetweenIterations (bool enable=true)

Static Public Attributes
static constexpr int64_t	kNoLimit = -1

Detailed Description

This class allows control over how the GreedyPatternRewriteDriver works.

Definition at line 43 of file GreedyPatternRewriteDriver.h.

Member Function Documentation

◆ enableConstantCSE()

GreedyRewriteConfig & mlir::GreedyRewriteConfig::enableConstantCSE ( bool enable = true )

inline

Definition at line 136 of file GreedyPatternRewriteDriver.h.

Referenced by mlirGreedyRewriteDriverConfigEnableConstantCSE().

◆ enableCSEBetweenIterations()

GreedyRewriteConfig & mlir::GreedyRewriteConfig::enableCSEBetweenIterations ( bool enable = true )

inline

Definition at line 159 of file GreedyPatternRewriteDriver.h.

◆ enableFolding()

GreedyRewriteConfig & mlir::GreedyRewriteConfig::enableFolding ( bool enable = true )

inline

Definition at line 128 of file GreedyPatternRewriteDriver.h.

Referenced by mlirGreedyRewriteDriverConfigEnableFolding().

◆ getListener()

RewriterBase::Listener * mlir::GreedyRewriteConfig::getListener ( ) const

inline

An optional listener that should be notified about IR modifications.

Definition at line 120 of file GreedyPatternRewriteDriver.h.

◆ getMaxIterations()

int64_t mlir::GreedyRewriteConfig::getMaxIterations ( ) const

inline

This specifies the maximum number of times the rewriter will iterate between applying patterns and simplifying regions.

Use kNoLimit to disable this iteration limit.

Note: Only applicable when simplifying entire regions.

Definition at line 76 of file GreedyPatternRewriteDriver.h.

Referenced by mlir::applyPatternsGreedily(), and mlirGreedyRewriteDriverConfigGetMaxIterations().

◆ getMaxNumRewrites()

int64_t mlir::GreedyRewriteConfig::getMaxNumRewrites ( ) const

inline

This specifies the maximum number of rewrites within an iteration.

Use kNoLimit to disable this limit.

Definition at line 84 of file GreedyPatternRewriteDriver.h.

Referenced by mlir::applyOpPatternsGreedily(), and mlirGreedyRewriteDriverConfigGetMaxNumRewrites().

◆ getRegionSimplificationLevel()

GreedySimplifyRegionLevel mlir::GreedyRewriteConfig::getRegionSimplificationLevel ( ) const

inline

Perform control flow optimizations to the region tree after applying all patterns.

Note: Only applicable when simplifying entire regions.

Definition at line 62 of file GreedyPatternRewriteDriver.h.

Referenced by mlirGreedyRewriteDriverConfigGetRegionSimplificationLevel().

◆ getScope()

Region * mlir::GreedyRewriteConfig::getScope ( ) const

inline

Only ops within the scope are added to the worklist.

If no scope is specified, the closest enclosing region around the initial list of ops (or the specified region, depending on which greedy rewrite entry point is used) is used as a scope. Any region being simplified must be enclosed by the scope (or equal to it).

Definition at line 97 of file GreedyPatternRewriteDriver.h.

Referenced by mlir::applyOpPatternsGreedily(), and mlir::applyPatternsGreedily().

◆ getStrictness()

GreedyRewriteStrictness mlir::GreedyRewriteConfig::getStrictness ( ) const

inline

Strict mode can restrict the ops that are added to the worklist during the rewrite.

GreedyRewriteStrictness::AnyOp: No ops are excluded.
GreedyRewriteStrictness::ExistingAndNewOps: Only pre-existing ops (that were on the worklist at the very beginning) and newly created ops are enqueued. All other ops are excluded.
GreedyRewriteStrictness::ExistingOps: Only pre-existing ops (that were were on the worklist at the very beginning) enqueued. All other ops are excluded.

Definition at line 113 of file GreedyPatternRewriteDriver.h.

Referenced by mlirGreedyRewriteDriverConfigGetStrictness().

◆ getUseTopDownTraversal()

bool mlir::GreedyRewriteConfig::getUseTopDownTraversal ( ) const

inline

This specifies the order of initial traversal that populates the rewriters worklist.

When set to true, it walks the operations top-down, which is generally more efficient in compile time. When set to false, its initial traversal of the region tree is bottom up on each block, which may match larger patterns when given an ambiguous pattern set.

Note: Only applicable when simplifying entire regions.

Definition at line 52 of file GreedyPatternRewriteDriver.h.

Referenced by mlirGreedyRewriteDriverConfigGetUseTopDownTraversal().

◆ isConstantCSEEnabled()

bool mlir::GreedyRewriteConfig::isConstantCSEEnabled ( ) const

inline

If set to "true", constants are CSE'd (even across multiple regions that are in a parent-ancestor relationship).

Definition at line 135 of file GreedyPatternRewriteDriver.h.

Referenced by mlirGreedyRewriteDriverConfigIsConstantCSEEnabled().

◆ isCSEBetweenIterationsEnabled()

bool mlir::GreedyRewriteConfig::isCSEBetweenIterationsEnabled ( ) const

inline

If set to "true", full common-subexpression elimination is run on the scoped region between each pattern-application iteration.

Unlike cseConstants (which only deduplicates constant ops via the operation folder) this runs the standard CSE algorithm and can unblock further canonicalizations on the next iteration. Off by default because it rebuilds dominance info each iteration.

Caveats when enabling this option:

Any listener attached via setListener will be notified of notifyOperationReplaced / notifyOperationErased events generated by CSE. Pattern authors relying on operation identity (e.g., the transform dialect's handle tracking) must account for this.
CSE-driven changes feed back into the iteration loop: a pattern that re-materializes duplicates that CSE keeps collapsing can extend the iteration count and, in the worst case, hit maxIterations. Under testConvergence=true such pipelines will be reported as non-convergent.

Definition at line 158 of file GreedyPatternRewriteDriver.h.