[mlir][shape] Generalize broadcast to a variadic number of shapes

[tpopp]

Summary

Previously broadcast was a binary op. Now it can support more inputs.
This has been changed in such a way that for now, this is an NFC for
all broadcast operations that were previously legal.

[tpopp]

Follow naming convention in new functions.

[jpienaar]

mlir/include/mlir/Dialect/Shape/IR/ShapeOps.td

83

Do we have places that use this form?

mlir/lib/Conversion/ShapeToStandard/ShapeToStandard.cpp

80

The last sentence is true for all rewrites, so redundant here.

90

A comment here would be good to explain why dim is used to get rank.

Also does dim work on a shape? (Tensor yes, and so are we guaranteed we'd be in tensor world here?)

188

Could more be reused here? E.g., is the nary one (excluding computing the max rank) less efficient for the binary case? Or vice versa, is multiple binary case applications less efficient than nary lowering?

[herhut]

mlir/lib/Conversion/ShapeToStandard/ShapeToStandard.cpp

115

inBound is confusing. It is true if we are outside the bounds of the index, right?

130

This sentence ends abruptly.

142

You could fold this up. In the then case, return reduceDim. In the else case, do the select.

188

Thinking of it, maybe doing multiple 2d broadcasts in a row (in the implementation, not the op) would yield similar performance.

mlir/lib/Dialect/Shape/IR/Shape.cpp

355

Is this intended?

[tpopp]

Fix variable names and adress other comments.

[tpopp]

I'm going to wait on fixing the test until I hear your thoughts on the two options for lowering broadcast.

mlir/include/mlir/Dialect/Shape/IR/ShapeOps.td

83

This was the original form of the build method before making the number of inputs variadic. I thought it might be nice to still have it for the common case and to make the transition essentially an NFC

mlir/lib/Conversion/ShapeToStandard/ShapeToStandard.cpp

90

This was guaranteed in the caller. I'll replicate the check here though in case of future uses from other locations.

188

I think performance differences between the two are negligible. We already fully unroll the tensor::GenerateOp, so they will be roughly the same.

For the binary case, we need a starting value during reductions in the nary lowering (without making the c++ code much more complex) while the binary case can skip that step. On the other hand, the binary case might recompute a small amount of work between each invocation for more than 2 inputs. Technically the nary case has the potential to be just as performant. I'm just not sure how clean I can make it look.

I personally find the n-ary lowering to be easier to read, and I think Stephan has in the past said he expected the binary case was easier to read. I think we should choose the implementation that we find easier to read and stick with that. I would like to hear your opinions on if you agree or not, and which you find easier to read.

[herhut]

mlir/lib/Conversion/ShapeToStandard/ShapeToStandard.cpp

188

I found the binary case easier because my n-ary case looked like a mess. Looking at your code, this is much nicer.

So let's ship the n-ary case only. The performance difference should be negligible.

Could you also extend the cstr_broadcastable accordingly? They need to be in sync otherwise broadcast cannot really be used.

[tpopp]

Removed binary broadcast case and updated tests accordingly.

[tpopp]

mlir/lib/Conversion/ShapeToStandard/ShapeToStandard.cpp

188

Binary is removed. I'll extend cstr_broadcastable in a follow up CL.

[herhut]

Please address minor nits before landing.

mlir/lib/Conversion/ShapeToStandard/ShapeToStandard.cpp

151

This comment seems lonely.

237–246

Does it still make sense to have this in an extra function?

[tpopp]

Replace naryBroadcast helper with only a single dim calculation.

[tpopp]

mlir/lib/Conversion/ShapeToStandard/ShapeToStandard.cpp

151

Fixed

237–246

I refactored to have only part of this function separate as it's convenient or use in the broadcastable follow up

mlir/include/mlir/Dialect/Shape/IR/ShapeOps.td

let description = [{
     In case the resulting shape is undefined, i.e. if corresponding extents are
     different from each other but none is 1, the result is an error shape.
     Likewise error values are propagated if any of the operands holds an error
     value. If the result type is an extent tensor (and can therefore not hold
     the error value) the behavior may be undefined. The optional string
     attribute can be used to describe the error case.
   }];
 
-  let arguments = (ins Shape_ShapeOrExtentTensorType:$lhs,
-                       Shape_ShapeOrExtentTensorType:$rhs,
+  let arguments = (ins Variadic<Shape_ShapeOrExtentTensorType>:$shapes,
                        OptionalAttr<StrAttr>:$error);
   let results = (outs Shape_ShapeOrExtentTensorType:$result);
 
   let assemblyFormat = [{
-    $lhs `,` $rhs attr-dict `:` type($lhs) `,` type($rhs) `->` type($result)
+    $shapes attr-dict `:` type($shapes) `->` type($result)
   }];
 
-  let verifier = [{ return ::verifyShapeOrExtentTensorOp(*this); }];
-  let hasFolder = 1;
+  let builders = [OpBuilderDAG<(ins "::mlir::Type":$result,
+                                "::mlir::Value":$lhs, "::mlir::Value":$rhs,
+                                "/*optional*/ ::mlir::StringAttr":$error), [{
+      build($_builder, $_state, result, ::llvm::makeArrayRef({lhs, rhs}), error);
+    }]>
+  ];
 
-  let verifier = [{ return ::verifyShapeOrExtentTensorOp(*this); }];
+  let hasFolder = 1;
+  let verifier = [{
+    return success(succeeded(::verifyShapeOrExtentTensorOp(*this)) &&
+                   getNumOperands() >= 2);
+  }];
 }
 
 def Shape_ConstShapeOp : Shape_Op<"const_shape", [ConstantLike, NoSideEffect]> {
   let summary = "Creates a constant shape or extent tensor";
   let description = [{
     Creates a constant shape or extent tensor. The individual extents are given
     as the `shape` attribute. The number of these values equals the shape's
     rank.

mlir/lib/Conversion/ShapeToStandard/ShapeToStandard.cpp

 //===- ShapeToStandard.cpp - conversion from Shape to Standard dialect ----===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #include "mlir/Conversion/ShapeToStandard/ShapeToStandard.h"
 
 #include "../PassDetail.h"
 #include "mlir/Dialect/SCF/SCF.h"
 #include "mlir/Dialect/Shape/IR/Shape.h"
 #include "mlir/Dialect/StandardOps/IR/Ops.h"
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
 #include "mlir/IR/BlockAndValueMapping.h"
+#include "mlir/IR/ImplicitLocOpBuilder.h"
 #include "mlir/Transforms/DialectConversion.h"
+#include "llvm/ADT/STLExtras.h"
 
 using namespace mlir;
 using namespace mlir::shape;
 using namespace mlir::scf;
 
 /// Conversion patterns.
 namespace {
 class AnyOpConversion : public OpConversionPattern<AnyOp> {

 namespace {
 struct BroadcastOpConverter : public OpConversionPattern<BroadcastOp> {
   using OpConversionPattern<BroadcastOp>::OpConversionPattern;
 
   LogicalResult
   matchAndRewrite(BroadcastOp op, ArrayRef<Value> operands,
                   ConversionPatternRewriter &rewriter) const override;
 };
-} // namespace
 
-LogicalResult BroadcastOpConverter::matchAndRewrite(
-    BroadcastOp op, ArrayRef<Value> operands,
-    ConversionPatternRewriter &rewriter) const {
-  // For now, this lowering is only defined on `tensor<?xindex>` operands, not
-  // on shapes.
-  if (op.getType().isa<ShapeType>())
-    return failure();
+// The generic lowering of Broadcast given any number of inputs. It assumes all
+// verification was done before being called.
+Operation::result_range naryBroadcastLowering(BroadcastOp::Adaptor &op,
+                                              ImplicitLocOpBuilder lb) {
+  Value zero = lb.create<ConstantIndexOp>(0);
+  Value one = lb.create<ConstantIndexOp>(1);
+  Type indexTy = lb.getIndexType();
+
+  // Save all the ranks for bounds checking
+  SmallVector<Value> ranks, rankDiffs;
+  llvm::append_range(ranks, llvm::map_range(op.shapes(), [&](Value v) {
+                       return lb.create<DimOp>(v, zero);
+                     }));
+
+  // Find the maximum rank
+  Value maxRank = ranks.front();
+  for (Value v : llvm::drop_begin(ranks, 1)) {
+    Value rankIsGreater = lb.create<CmpIOp>(CmpIPredicate::ugt, v, maxRank);
+    maxRank = lb.create<SelectOp>(rankIsGreater, v, maxRank);
+  }
+
+  // Calculate the difference of ranks and the maximum rank for later offsets.
+  llvm::append_range(rankDiffs, llvm::map_range(ranks, [&](Value v) {
+                       return lb.create<SubIOp>(indexTy, maxRank, v);
+                     }));
+
+  return lb
+      .create<tensor::GenerateOp>(
+          getExtentTensorType(lb.getContext()), ValueRange{maxRank},
+          [&](OpBuilder &b, Location loc, ValueRange args) {
+            Value outputDimension = args[0];
 
-  assert(!op.lhs().getType().isa<ShapeType>() &&
-         !op.rhs().getType().isa<ShapeType>());
-  auto loc = op.getLoc();
-  BroadcastOp::Adaptor transformed(operands);
-  Value zero = rewriter.create<ConstantIndexOp>(loc, 0);
-  Value one = rewriter.create<ConstantIndexOp>(loc, 1);
+            Value reduceDim = one;
+            for (auto tup : llvm::zip(op.shapes(), rankDiffs)) {
+              Value shape = std::get<0>(tup);
+              Value rankDiff = std::get<1>(tup);
+              Value inBound = b.create<CmpIOp>(loc, CmpIPredicate::ult,
+                                               outputDimension, rankDiff);
+              Value dim =
+                  b.create<IfOp>(
+                       loc, TypeRange{indexTy}, inBound,
+                       [&](OpBuilder &b, Location loc) {
+                         b.create<scf::YieldOp>(loc, one);
+                       },
+                       [&](OpBuilder &b, Location loc) {
+                         // The broadcasting logic is:
+                         // - if one extent (here we arbitrarily choose the
+                         // extent from the greater-rank operand) is equal to 1,
+                         // then take the extent from the other operand
+                         // - otherwise, take the extent as-is.
+                         // Note that this logic remains correct in the presence
+                         // of
+                         Value lesserRankOperandDimension = b.create<SubIOp>(
+                             loc, indexTy, outputDimension, rankDiff);
+                         Value lesserRankOperandExtent =
+                             b.create<tensor::ExtractOp>(
+                                 loc, shape,
+                                 ValueRange{lesserRankOperandDimension});
+                         b.create<scf::YieldOp>(loc, lesserRankOperandExtent);
+                       })
+                      .getResult(0);
+
+              // Always give preference to a possibly non-1 extent
+              Value dimIsOne =
+                  b.create<CmpIOp>(loc, CmpIPredicate::eq, dim, one);
+              reduceDim = b.create<SelectOp>(loc, dimIsOne, reduceDim, dim);
+            }
+
+            b.create<tensor::YieldOp>(loc, reduceDim);
+          })
+      ->getResults();
+}
+
+// The specialized lowering for the common case binary broadcast case. This is
+// slightly more efficient and and arguably easier to read. It assumes all
+// verification was done before being called.
+Operation::result_range binaryBroadcastLowering(BroadcastOp::Adaptor &op,
+                                                ImplicitLocOpBuilder lb) {
+  Value zero = lb.create<ConstantIndexOp>(0);
+  Value one = lb.create<ConstantIndexOp>(1);
+  Type indexTy = lb.getIndexType();
 
   // Find smaller and greater rank and extent tensor.
-  Value lhsRank = rewriter.create<DimOp>(loc, op.lhs(), zero);
-  Value rhsRank = rewriter.create<DimOp>(loc, op.rhs(), zero);
-  Value lhsRankULE =
-      rewriter.create<CmpIOp>(loc, CmpIPredicate::ule, lhsRank, rhsRank);
-  Type indexTy = rewriter.getIndexType();
-  Value lesserRank =
-      rewriter.create<SelectOp>(loc, lhsRankULE, lhsRank, rhsRank);
-  Value greaterRank =
-      rewriter.create<SelectOp>(loc, lhsRankULE, rhsRank, lhsRank);
+  Value lhsRank = lb.create<DimOp>(op.shapes()[0], zero);
+  Value rhsRank = lb.create<DimOp>(op.shapes()[1], zero);
+  Value lhsRankULE = lb.create<CmpIOp>(CmpIPredicate::ule, lhsRank, rhsRank);
+  Value lesserRank = lb.create<SelectOp>(lhsRankULE, lhsRank, rhsRank);
+  Value greaterRank = lb.create<SelectOp>(lhsRankULE, rhsRank, lhsRank);
   auto erasedRankType =
       RankedTensorType::get({ShapedType::kDynamicSize}, indexTy);
   Value rankErasedLhs =
-      rewriter.create<tensor::CastOp>(loc, erasedRankType, transformed.lhs());
+      lb.create<tensor::CastOp>(erasedRankType, op.shapes()[0]);
   Value rankErasedRhs =
-      rewriter.create<tensor::CastOp>(loc, erasedRankType, transformed.rhs());
+      lb.create<tensor::CastOp>(erasedRankType, op.shapes()[1]);
   Value lesserRankOperand =
-      rewriter.create<SelectOp>(loc, lhsRankULE, rankErasedLhs, rankErasedRhs);
+      lb.create<SelectOp>(lhsRankULE, rankErasedLhs, rankErasedRhs);
   Value greaterRankOperand =
-      rewriter.create<SelectOp>(loc, lhsRankULE, rankErasedRhs, rankErasedLhs);
+      lb.create<SelectOp>(lhsRankULE, rankErasedRhs, rankErasedLhs);
 
-  Value rankDiff =
-      rewriter.create<SubIOp>(loc, indexTy, greaterRank, lesserRank);
-  rewriter.replaceOpWithNewOp<tensor::GenerateOp>(
-      op, getExtentTensorType(op.getContext()), ValueRange{greaterRank},
+  Value rankDiff = lb.create<SubIOp>(indexTy, greaterRank, lesserRank);
+  return lb
+      .create<tensor::GenerateOp>(
+          getExtentTensorType(lb.getContext()), ValueRange{greaterRank},
           [&](OpBuilder &b, Location loc, ValueRange args) {
             Value outputDimension = args[0];
             Value isUnchallengedDimension = b.create<CmpIOp>(
                 loc, CmpIPredicate::ult, outputDimension, rankDiff);
             Value greaterRankOperandExtent = b.create<tensor::ExtractOp>(
                 loc, greaterRankOperand, outputDimension);
-        // The initial dimensions of the greater-rank operand are unchallenged,
-        // so we can take them as-is. Otherwise, we need to do a comparison.
-        // We need an actual branch here (instead of a select) because the
-        // lesser-rank operand might be rank 0, so any tensor.extract would be
-        // invalid.
+            // The initial dimensions of the greater-rank operand are
+            // unchallenged, so we can take them as-is. Otherwise, we need to do
+            // a comparison. We need an actual branch here (instead of a select)
+            // because the lesser-rank operand might be rank 0, so any
+            // tensor.extract would be invalid.
             auto ifOp = b.create<IfOp>(
                 loc, TypeRange{indexTy}, isUnchallengedDimension,
                 [&](OpBuilder &b, Location loc) {
                   b.create<scf::YieldOp>(loc, greaterRankOperandExtent);
                 },
                 [&](OpBuilder &b, Location loc) {
                   // The broadcasting logic is:
                   // - if one extent (here we arbitrarily choose the extent from
-              // the greater-rank operand) is equal to 1, then take the extent
-              // from the other operand
+                  // the greater-rank operand) is equal to 1, then take the
+                  // extent from the other operand
                   // - otherwise, take the extent as-is.
                   // Note that this logic remains correct in the presence of
                   // dimensions of zero extent.
                   Value lesserRankOperandDimension =
                       b.create<SubIOp>(loc, indexTy, outputDimension, rankDiff);
                   Value lesserRankOperandExtent = b.create<tensor::ExtractOp>(
                       loc, lesserRankOperand,
                       ValueRange{lesserRankOperandDimension});
                   Value greaterRankOperandExtentIsOne = b.create<CmpIOp>(
                       loc, CmpIPredicate::eq, greaterRankOperandExtent, one);
                   Value broadcastedExtent = b.create<SelectOp>(
-                  loc, greaterRankOperandExtentIsOne, lesserRankOperandExtent,
-                  greaterRankOperandExtent);
+                      loc, greaterRankOperandExtentIsOne,
+                      lesserRankOperandExtent, greaterRankOperandExtent);
                   b.create<scf::YieldOp>(loc, broadcastedExtent);
                 });
             b.create<tensor::YieldOp>(loc, ifOp.getResult(0));
-      });
+          })
+      ->getResults();
+}
+
+} // namespace
+
+LogicalResult BroadcastOpConverter::matchAndRewrite(
+    BroadcastOp op, ArrayRef<Value> operands,
+    ConversionPatternRewriter &rewriter) const {
+  // For now, this lowering is only defined on `tensor<?xindex>` operands, not
+  // on shapes.
+  if (op.getType().isa<ShapeType>())
+    return failure();
+
+  BroadcastOp::Adaptor transformed(operands);
+  assert(llvm::all_of(transformed.shapes(),
+                      [](Value v) { return !v.getType().isa<ShapeType>(); }));
+
+  auto loc = op.getLoc();
+  ImplicitLocOpBuilder lb(loc, rewriter);
+
+  // Specialize case the 2 input case as it is slightly optimized and might be
+  // slightly easier to read.
+  if (transformed.shapes().size() == 2) {
+    rewriter.replaceOp(op, binaryBroadcastLowering(transformed, lb));
+  } else {
+    rewriter.replaceOp(op, naryBroadcastLowering(transformed, lb));
+  }
   return success();
 }
 
 namespace {
 class ConstShapeOpConverter : public OpConversionPattern<ConstShapeOp> {
 public:
   using OpConversionPattern<ConstShapeOp>::OpConversionPattern;
 

mlir/lib/Dialect/Shape/IR/Shape.cpp

if (op.getNumOperands() == 0)
     return op.emitOpError("no operands specified");
 
   return success();
 }
 
 //===----------------------------------------------------------------------===//
 // BroadcastOp
 //===----------------------------------------------------------------------===//
-
+//
 OpFoldResult BroadcastOp::fold(ArrayRef<Attribute> operands) {
   if (!operands[1])
     return nullptr;
 
+  // TODO: Support folding with more than 2 input shapes
+  if (operands.size() > 2 && !operands[2].isa<StringAttr>())
+    return nullptr;
+
   auto rhsShape = llvm::to_vector<6>(
       operands[1].cast<DenseIntElementsAttr>().getValues<int64_t>());
   if (rhsShape.empty())
-    return lhs();
+    return shapes()[0];
 
   if (!operands[0])
     return nullptr;
 
   auto lhsShape = llvm::to_vector<6>(
       operands[0].cast<DenseIntElementsAttr>().getValues<int64_t>());
   if (lhsShape.empty())
-    return rhs();
+    return shapes()[1];
 
   SmallVector<int64_t, 6> resultShape;
   // If the shapes are not compatible, we can't fold it.
   // TODO: Fold to an "error".
   if (!OpTrait::util::getBroadcastedShape(lhsShape, rhsShape, resultShape))
     return nullptr;
   Builder builder(getContext());
   return builder.getIndexTensorAttr(resultShape);

mlir/test/Conversion/ShapeToStandard/shape-to-standard.mlir

 // CHECK:             %[[EITHER_EXTENT_IS_ONE:.*]] = or %[[LARGER_EXTENT_IS_ONE]], %[[SMALLER_EXTENT_IS_ONE]] : i1
 // CHECK:             %[[OR_EXTENTS_ARE_EQUAL:.*]] = or %[[EITHER_EXTENT_IS_ONE]], %[[EXTENTS_ARE_EQUAL]] : i1
 // CHECK:             %[[NEW_ALL_SO_FAR:.*]] = and %[[ALL_SO_FAR]], %[[OR_EXTENTS_ARE_EQUAL]] : i1
 // CHECK:             scf.yield %[[NEW_ALL_SO_FAR]] : i1
 // CHECK:           }
 // CHECK:           %[[RESULT:.*]] = shape.cstr_require %[[ALL_RESULT]], "required broadcastable shapes"
 // CHECK:           return %[[RESULT]] : !shape.witness
 // CHECK:         }
+
+// -----
+
+func @broadcast_3_shapes_different_extents(%a : tensor<2xindex>,
+                                           %b : tensor<3xindex>,
+                                           %c : tensor<2xindex>) {
+// CHECK-LABEL:   func @broadcast_3_shapes_different_extents(
+// CHECK-SAME:          %[[ARG0:.*]]: tensor<2xindex>,
+// CHECK-SAME:          %[[ARG1:.*]]: tensor<3xindex>,
+// CHECK-SAME:          %[[ARG2:.*]]: tensor<2xindex>) {
+// CHECK:           %[[C0:.*]] = constant 0 : index
+// CHECK:           %[[C1:.*]] = constant 1 : index
+// CHECK:           %[[RANK0:.*]] = dim %[[ARG0]], %[[C0]] : tensor<2xindex>
+// CHECK:           %[[RANK1:.*]] = dim %[[ARG1]], %[[C0]] : tensor<3xindex>
+// CHECK:           %[[RANK2:.*]] = dim %[[ARG2]], %[[C0]] : tensor<2xindex>
+// CHECK:           %[[CMP0:.*]] = cmpi ugt, %[[RANK1]], %[[RANK0]] : index
+// CHECK:           %[[LARGER_DIM:.*]] = select %[[CMP0]], %[[RANK1]], %[[RANK0]] : index
+// CHECK:           %[[CMP1:.*]] = cmpi ugt, %[[RANK2]], %[[LARGER_DIM]] : index
+// CHECK:           %[[MAX_RANK:.*]] = select %[[CMP1]], %[[RANK2]], %[[LARGER_DIM]] : index
+// CHECK:           %[[DIM_DIFF0:.*]] = subi %[[MAX_RANK]], %[[RANK0]] : index
+// CHECK:           %[[DIM_DIFF1:.*]] = subi %[[MAX_RANK]], %[[RANK1]] : index
+// CHECK:           %[[DIM_DIFF2:.*]] = subi %[[MAX_RANK]], %[[RANK2]] : index
+// CHECK:           %[[RESULT:.*]] = tensor.generate %[[MAX_RANK]]  {
+// CHECK:           ^bb0(%[[IDX:.*]]: index):
+// CHECK:             %[[OUTBOUNDS0:.*]] = cmpi ult, %[[IDX]], %[[DIM_DIFF0]] : index
+// CHECK:             %[[DIM0:.*]] = scf.if %[[OUTBOUNDS0]] -> (index) {
+// CHECK:               scf.yield %[[C1]] : index
+// CHECK:             } else {
+// CHECK:               %[[IDX0:.*]] = subi %[[IDX]], %[[DIM_DIFF0]] : index
+// CHECK:               %[[VAL_23:.*]] = tensor.extract %[[ARG0]]{{\[}}%[[IDX0]]] : tensor<2xindex>
+// CHECK:               scf.yield %[[VAL_23]] : index
+// CHECK:             }
+// CHECK:             %[[DIM0_IS_1:.*]] = cmpi eq, %[[DIM0:.*]], %[[C1]] : index
+// CHECK:             %[[MAX_DIM0:.*]] = select %[[DIM0_IS_1]], %[[C1]], %[[DIM0]] : index
+// CHECK:             %[[VAL_28:.*]] = cmpi ult, %[[IDX]], %[[DIM_DIFF1]] : index
+// CHECK:             %[[DIM1:.*]] = scf.if %[[VAL_28]] -> (index) {
+// CHECK:               scf.yield %[[C1]] : index
+// CHECK:             } else {
+// CHECK:               %[[IDX1:.*]] = subi %[[IDX]], %[[DIM_DIFF1]] : index
+// CHECK:               %[[VAL_31:.*]] = tensor.extract %[[ARG1]]{{\[}}%[[IDX1]]] : tensor<3xindex>
+// CHECK:               scf.yield %[[VAL_31]] : index
+// CHECK:             }
+// CHECK:             %[[DIM1_IS_1:.*]] = cmpi eq, %[[DIM1:.*]], %[[C1]] : index
+// CHECK:             %[[MAX_DIM1:.*]] = select %[[DIM1_IS_1]], %[[MAX_DIM0]], %[[DIM1]] : index
+// CHECK:             %[[VAL_36:.*]] = cmpi ult, %[[IDX]], %[[DIM_DIFF2]] : index
+// CHECK:             %[[DIM2:.*]] = scf.if %[[VAL_36]] -> (index) {
+// CHECK:               scf.yield %[[C1]] : index
+// CHECK:             } else {
+// CHECK:               %[[IDX2:.*]] = subi %[[IDX]], %[[DIM_DIFF2]] : index
+// CHECK:               %[[VAL_39:.*]] = tensor.extract %[[ARG2]]{{\[}}%[[IDX2]]] : tensor<2xindex>
+// CHECK:               scf.yield %[[VAL_39]] : index
+// CHECK:             }
+// CHECK:             %[[DIM2_IS_1:.*]] = cmpi eq, %[[DIM2:.*]], %[[C1]] : index
+// CHECK:             %[[MAX_DIM2:.*]] = select %[[DIM2_IS_1]], %[[MAX_DIM1]], %[[DIM2]] : index
+// CHECK:             tensor.yield %[[MAX_DIM2]] : index
+// CHECK:           } : tensor<?xindex>
+// CHECK:           return
+// CHECK:         }
+  %0 = shape.broadcast %a, %b, %c
+      : tensor<2xindex>, tensor<3xindex>, tensor<2xindex> -> tensor<?xindex>
+  return
+}