[compiler] refactor anonymous patterns to named patterns (#81)

* [compiler] refactor anonymous patterns to named patterns * update
bytedance · Nov 15, 2023 · 06b18bf · 06b18bf
1 parent 16da0f9
commit 06b18bf
Show file tree

Hide file tree

Showing 3 changed files with 980 additions and 964 deletions.
diff --git a/compiler/include/byteir/Dialect/mhlo/Transforms/CanonicalizeExt.h b/compiler/include/byteir/Dialect/mhlo/Transforms/CanonicalizeExt.h
@@ -24,122 +24,13 @@
 namespace mlir {
 class MLIRContext;
 
-namespace tensor {
-class InsertSliceOp;
-}
-
 namespace mhlo {
-class AddOp;
-class ClampOp;
-class ConvertOp;
-class CompareOp;
-class CustomCallOp;
-class TransposeOp;
-class BroadcastInDimOp;
-class ConcatenateOp;
-class DynamicBroadcastInDimOp;
-class DynamicConvOp;
-class DynamicGatherOp;
-class ReduceWindowOp;
-class ReshapeOp;
-class MulOp;
-class SliceOp;
-class ReverseOp;
-class GatherOp;
 
 // Most of these will push back to upstream
 // So this file only includes patterns, not a pass.
 
-///
-///  foldBroadcastInDimConstWithBinary
-///
-/// BroadcastInDim could be folded in some special cases. Ex.
-///
-/// const
-///   \
-///   broadcast_in_dim  const
-///       \              /
-///             mul
-LogicalResult foldBroadcastInDimConstWithBinary(mhlo::BroadcastInDimOp op,
-                                                PatternRewriter &rewriter);
-
-///
-///  Fold concatenate of continuous slices
-///
-LogicalResult foldConcatWithContinuousSlices(mhlo::ConcatenateOp op,
-                                             PatternRewriter &rewriter);
-
 // fold multi op with zero
-LogicalResult foldMultiplyZero(mhlo::MulOp op, PatternRewriter &rewriter);
-
-// fold binary op with large constant op
-template <typename Op, template <typename> typename Func>
-LogicalResult foldLargeBinaryOp(Op op, PatternRewriter &rewriter);
-
-// mhlo.dynamic_conv => mhlo.convolution canonicalization
-LogicalResult simplifyDynamicConvToConv(mhlo::DynamicConvOp op,
-                                        PatternRewriter &rewriter);
-
-// constant folding for mhlo.concatenate with large result
-LogicalResult foldLargeConcatenate(mhlo::ConcatenateOp op,
-                                   PatternRewriter &rewriter);
-
-LogicalResult foldTransposeNonSplat(mhlo::TransposeOp op,
-                                    PatternRewriter &rewriter);
-
-LogicalResult foldBeneficialConstantConvertOp(mhlo::ConvertOp op,
-                                              PatternRewriter &rewriter);
-
-LogicalResult foldLargeCompareOp(mhlo::CompareOp op, PatternRewriter &rewriter);
-
-// const + broadcast_in_dim => const + broadcast_in_dim
-LogicalResult canonicalizeBroadcastInDimConst(mhlo::BroadcastInDimOp op,
-                                              PatternRewriter &rewriter);
-
-// simplify an addOp of two chain of insert_slice's
-// into a chain of insert_slice's
-// when those insert_slice's are
-// 1) not overlaped
-// 2) along a single axis
-// 3) sharing a zero Dest
-LogicalResult simplifyAddInsertSlicesToInsertSlices(mhlo::AddOp op,
-                                                    PatternRewriter &rewriter);
-
-// simplify a chain of insert_slice's into a concat
-// when those insert_slice's are
-// 1) not overlaped
-// 2) along a single axis
-// 3) covering the entire Dest
-LogicalResult simplifyFullInsertSlicesToConcat(mlir::tensor::InsertSliceOp op,
-                                               PatternRewriter &rewriter);
-
-// simplify byteir.addn => mhlo.add
-LogicalResult simplifyByteIRAddNToAdd(mhlo::CustomCallOp op,
-                                      PatternRewriter &rewriter);
-
-LogicalResult foldLargeSliceOp(mhlo::SliceOp op, PatternRewriter &rewriter);
-
-LogicalResult foldConcatWithSlicesAndRehape(mhlo::ConcatenateOp op,
-                                            PatternRewriter &rewriter);
-
-// concat(broadcast_in_dim(x), broadcast_in_dim(x)) => broadcast_in_dim
-LogicalResult canonicalizeConcatWithBroadcast(mhlo::ConcatenateOp op,
-                                              PatternRewriter &rewriter);
-
-LogicalResult eliminateRedundantConvertFromI1(mhlo::ConvertOp op,
-                                              PatternRewriter &rewriter);
-
-LogicalResult simplifyCumsumToIota(mhlo::ReduceWindowOp op,
-                                   PatternRewriter &rewriter);
-
-// transpose(reshape(transpose(x))) => reshape(x)
-LogicalResult simplifyTransposeReshapeTranspose(mhlo::TransposeOp op,
-                                                PatternRewriter &rewriter);
-
-LogicalResult foldReverseWithConstant(mhlo::ReverseOp op,
-                                      PatternRewriter &rewriter);
-
-LogicalResult foldGatherWithInput(mhlo::GatherOp op, PatternRewriter &rewriter);
+void populateFoldMultiplyZeroPattern(RewritePatternSet &patterns);
 
 // populate canonicalizeExt patterns
 void populateCanonicalizeExtPatterns(RewritePatternSet &patterns,