[Mlir-commits] [mlir] [MLIR] Remove redundant transform passes dialect dependencies (PR #213022)

[ioana ghiban]

ioghiban wrote:

List of passes where self-dialect dependencies were removed, after testing:
Affine:
- SimplifyAffineWithBounds - matches Affine ops

ArmSME:
- OuterProductFusion - matches ArmSME ops.

Async:
- AsyncToAsyncRuntimePass, AsyncFuncToAsyncRuntimePass - match Async ops.
- AsyncRuntimeRefCountingPass, AsyncRuntimeRefCountingOptPass, AsyncRuntimePolicyBasedRefCountingPass - match Async ops or values with Async types.

Bufferization: 
- BufferDeallocationSimplificationPass - matches Bufferization ops.

EmitC:
- WrapFuncInClassPass, MLGOAddReflectionMapPass - match EmitC ops.
- FormExpressionsPass - kept TableGen dependency, removed C++ override that also registered EmitC

GPU:
- GpuLaunchSinkIndexComputationsPass, GpuEliminateBarriers, GpuDecomposeMemrefsPass - match GPU ops or require an enclosing `gpu.launch`.

Linalg:
- LinalgSpecializeGenericOpsPass, SimplifyDepthwiseConvPass, LinalgFoldUnitExtentDimsPass, LinalgBlockPackMatmul - match Linalg ops.
- ConvertLinalgToAffineLoopsPass, ConvertLinalgToLoopsPass, ConvertLinalgToParallelLoopsPass - match the LinalgOp interface but create no Linalg entities.

Math:
- MathExtendToSupportedTypes, MathSincosFusionPass - match Math ops.

OpenACC:
- ACCImplicitData, ACCImplicitDeclare, ACCImplicitRoutine, ACCDeclareGPUModuleInsertion, ACCLegalizeSerial, ACCEmitRemarksPrivate, ACCLoopTiling, ACCSpecializeForDevice, ACCSpecializeForHost, ACCIfClauseLowering, ACCRecipeMaterialization, ACCComputeLowering, ACCRoutineLowering, ACCCGToGPU - match OpenACC ops/interfaces or require OpenACC typed attributes.

OpenMP: 
- MarkDeclareTargetPass - requires an existing OpenMP target op or declare-target attribute.

Shard:
- ShardingPropagation, ShardSimplify, Partition - match Shard ops. Partition also redundantly inserts ShardDialect in its C++ getDependentDialects.

SparseTensor: 
- SparseAssembler - matches functions carrying SparseTensor encoding attributes.
- SparseReinterpretMap, PreSparsificationRewrite, LowerSparseOpsToForeach - match SparseTensor ops or ops using sparse-encoded types.

Tensor: 
- FoldTensorSubsetOpsPass - matches Tensor ops.

TOSA: 
- TosaInferShapesPass, TosaNarrowI64ToI32Pass, TosaNarrowF64ToF32Pass - match TOSA ops.

Quant:
- LowerQuantOps, NormalizeQuantTypes, StripFuncQuantTypes - match Quant ops or Quant types.

XeGPU:
- XeGPUPropagateLayout, XeGPUWgToSgDistribute, XeGPUBlocking, XeGPUPeepHoleOptimizer, XeGPUSgToLaneDistribute - match XeGPU ops, types, or layout attributes.

https://github.com/llvm/llvm-project/pull/213022