[MLIR][HLO] Eliminate duplicate broadcastable constraints

PiperOrigin-RevId: 376718433

lib/Dialect/mhlo/transforms/move_up_dynamic_broadcasts_for_fusion.cc

@@ -306,25 +306,58 @@ struct MergeAssumingOpsPattern : public OpRewritePattern<shape::AssumingOp> {
   }
 };
 
-struct EliminateDuplicateCstrBroadcastableOps
+// Eliminate casted extent tensors. Instead, produce the concrete extent tensor
-    : public OpRewritePattern<shape::CstrBroadcastableOp> {
+// type where possible.
-  using OpRewritePattern<shape::CstrBroadcastableOp>::OpRewritePattern;
+struct CanonicalizeCastedShapeOfOpPattern
+    : public OpRewritePattern<tensor::CastOp> {
+  using OpRewritePattern<tensor::CastOp>::OpRewritePattern;
 
-  LogicalResult matchAndRewrite(shape::CstrBroadcastableOp op,
+  LogicalResult matchAndRewrite(tensor::CastOp op,
                                 PatternRewriter &rewriter) const override {
-    // Search for previous occurence of the same constraint.
+    // Only merge tensor cast into `shape_of` ops.
-    Operation *it = op->getPrevNode();
+    auto shape_of_op = op.source().getDefiningOp<shape::ShapeOfOp>();
-    while (it != nullptr) {
+    if (!shape_of_op) return failure();
-      if (auto candidate = llvm::dyn_cast<shape::CstrBroadcastableOp>(it)) {
+
-        if (candidate.shapes() == op.shapes()) {
+    // Desired type must be an extent tensor type.
-          rewriter.replaceOp(op, candidate.result());
+    auto result_ty = op.getType().dyn_cast<RankedTensorType>();
+    if (!result_ty || result_ty.getRank() != 1 ||
+        !result_ty.getElementType().isIndex())
+      return failure();
+
+    rewriter.replaceOpWithNewOp<shape::ShapeOfOp>(op, result_ty,
+                                                  shape_of_op.arg());
+    if (shape_of_op->getUses().empty()) rewriter.eraseOp(shape_of_op);
     return success();
   }
+};
+
+// TODO(frgossen): Remove this once it has landed upstream.
+struct CanonicalizeBroadcastPattern
+    : public OpRewritePattern<shape::BroadcastOp> {
+  using OpRewritePattern<shape::BroadcastOp>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(shape::BroadcastOp op,
+                                PatternRewriter &rewriter) const override {
+    // Only concretize dynamic extent tensor result types.
+    auto resultTy = op.getType().dyn_cast<RankedTensorType>();
+    if (!resultTy || !resultTy.isDynamicDim(0)) return failure();
+
+    // Infer resulting shape rank if possible.
+    int64_t maxRank = 0;
+    for (Value shape : op.shapes()) {
+      if (auto extentTensorTy = shape.getType().dyn_cast<RankedTensorType>()) {
+        // Cannot infer resulting shape rank if any operand is dynamically
+        // ranked.
+        if (extentTensorTy.isDynamicDim(0)) return failure();
+        maxRank = std::max(maxRank, extentTensorTy.getDimSize(0));
       }
-      it = it->getPrevNode();
     }
 
-    return failure();
+    auto newOp = rewriter.create<shape::BroadcastOp>(
+        op.getLoc(), RankedTensorType::get({maxRank}, rewriter.getIndexType()),
+        op.shapes());
+    rewriter.replaceOpWithNewOp<tensor::CastOp>(op, op.getType(), newOp);
+    return success();
   }
 };
 
@@ -399,7 +432,8 @@ void PopulateMoveUpDynamicBroadcastsForFusionPatterns(
     MLIRContext *context, OwningRewritePatternList *patterns) {
   // clang-format off
   patterns->insert<
-      EliminateDuplicateCstrBroadcastableOps,
+      CanonicalizeBroadcastPattern,
+      CanonicalizeCastedShapeOfOpPattern,
       InlineBroadcastedShapeOperandsPattern<shape::CstrBroadcastableOp>,
       MergeAssumingOpsPattern,
       MoveIntoAssumingOpPattern<shape::ShapeOfOp>,
@@ -409,7 +443,6 @@ void PopulateMoveUpDynamicBroadcastsForFusionPatterns(
       MoveUpBroadcastInDimOpPattern,
       ShapeReificationPattern>(context);
   // clang-format on
-  shape::BroadcastOp::getCanonicalizationPatterns(*patterns, context);
   tensor::CastOp::getCanonicalizationPatterns(*patterns, context);
 }
 

tests/move_up_dynamic_broadcasts_for_fusion.mlir

@@ -374,21 +374,3 @@ func @sub_sub(%arg0: tensor<?x32xf16>, %arg1 : tensor<?x32xf16>,
   }
   return %7 : tensor<?x?x32xf16>
 }
-
-// -----
-
-// CHECK-LABEL: @redundant_cstr_broadcastable
-// CHECK-SAME: (%[[ARG0:.*]]: tensor<?xindex>, %[[ARG1:.*]]: tensor<?xindex>)
-func @redundant_cstr_broadcastable(%arg0: tensor<?xindex>,
-    %arg1 : tensor<?xindex>) {
-  // CHECK-DAG:  %[[WITNESS:.*]] = shape.cstr_broadcastable %[[ARG0]], %[[ARG1]]
-  // CHECK:      shape.assuming %[[WITNESS]]
-  %0 = shape.cstr_broadcastable %arg0, %arg1 : tensor<?xindex>, tensor<?xindex>
-  %1 = shape.cstr_broadcastable %arg0, %arg1 : tensor<?xindex>, tensor<?xindex>
-  %2 = shape.assuming_all %0, %1
-  shape.assuming %2 -> () {
-    "some.op"() : () -> ()
-    shape.assuming_yield
-  }
-  return
-}