[mlir][vector] Make the in_bounds attribute mandatory (#97049)

At the moment, the in_bounds attribute has two confusing/contradicting
properties:
  1. It is both optional _and_ has an effective default-value.
  2. The default value is "out-of-bounds" for non-broadcast dims, and
     "in-bounds" for broadcast dims.

(see the `isDimInBounds` vector interface method for an example of this
"default" behaviour [1]).

This PR aims to clarify the logic surrounding the `in_bounds` attribute
by:
  * making the attribute mandatory (i.e. it is always present),
  * always setting the default value to "out of bounds" (that's
    consistent with the current behaviour for the most common cases).

#### Broadcast dimensions in tests

As per [2], the broadcast dimensions requires the corresponding
`in_bounds` attribute to be `true`:
```
  vector.transfer_read op requires broadcast dimensions to be in-bounds
```

The changes in this PR mean that we can no longer rely on the
default value in cases like the following (dim 0 is a broadcast dim):
```mlir
  %read = vector.transfer_read %A[%base1, %base2], %f, %mask
      {permutation_map = affine_map<(d0, d1) -> (0, d1)>} :
    memref<?x?xf32>, vector<4x9xf32>
```

Instead, the broadcast dimension has to explicitly be marked as "in
bounds:

```mlir
  %read = vector.transfer_read %A[%base1, %base2], %f, %mask
      {in_bounds = [true, false], permutation_map = affine_map<(d0, d1) -> (0, d1)>} :
    memref<?x?xf32>, vector<4x9xf32>
```

All tests with broadcast dims are updated accordingly.

#### Changes in "SuperVectorize.cpp" and "Vectorization.cpp"

The following patterns in "Vectorization.cpp" are updated to explicitly
set the `in_bounds` attribute to `false`:
* `LinalgCopyVTRForwardingPattern` and `LinalgCopyVTWForwardingPattern`

Also, `vectorizeAffineLoad` (from "SuperVectorize.cpp") and
`vectorizeAsLinalgGeneric` (from "Vectorization.cpp") are updated to
make sure that xfer Ops created by these hooks set the dimension
corresponding to broadcast dims as "in bounds". Otherwise, the Op
verifier would complain

Note that there is no mechanism to verify whether the corresponding
memory access are indeed in bounds. Still, this is consistent with the
current behaviour where the broadcast dim would be implicitly assumed
to be "in bounds".

[1]
https://github.com/llvm/llvm-project/blob/4145ad2bac4bb99d5034d60c74bb2789f6c6e802/mlir/include/mlir/Interfaces/VectorInterfaces.td#L243-L246
[2]
https://mlir.llvm.org/docs/Dialects/Vector/#vectortransfer_read-vectortransferreadop

Author: banach-space

mlir/include/mlir/Dialect/Vector/IR/VectorOps.td

@@ -1363,7 +1363,7 @@ def Vector_TransferReadOp :
                    AffineMapAttr:$permutation_map,
                    AnyType:$padding,
                    Optional<VectorOf<[I1]>>:$mask,
-                   OptionalAttr<BoolArrayAttr>:$in_bounds)>,
+                   BoolArrayAttr:$in_bounds)>,
     Results<(outs AnyVectorOfAnyRank:$vector)> {
 
   let summary = "Reads a supervector from memory into an SSA vector value.";
@@ -1401,16 +1401,19 @@ def Vector_TransferReadOp :
     permutation or broadcasting. Elements whose corresponding mask element is
     `0` are masked out and replaced with `padding`.
 
-    An optional boolean array attribute `in_bounds` specifies for every vector
-    dimension if the transfer is guaranteed to be within the source bounds. If
-    specified, the `in_bounds` array length has to be equal to the vector rank.
-    If set to "false", accesses (including the starting point) may run
+    For every vector dimension, the boolean array attribute `in_bounds`
+    specifies if the transfer is guaranteed to be within the source bounds. If
+    set to "false", accesses (including the starting point) may run
     out-of-bounds along the respective vector dimension as the index increases.
-    Broadcast dimensions must always be in-bounds. In absence of the attribute,
-    accesses along all vector dimensions (except for broadcasts) may run
-    out-of-bounds. A `vector.transfer_read` can be lowered to a simple load if
-    all dimensions are specified to be within bounds and no `mask` was
-    specified. Note that non-vector dimensions *must* always be in-bounds.
+    Non-vector and broadcast dimensions *must* always be in-bounds. The
+    `in_bounds` array length has to be equal to the vector rank. This attribute
+    has a default value: `false` (i.e. "out-of-bounds"). When skipped in the
+    textual IR, the default value is assumed. Similarly, the OP printer will
+    omit this attribute when all dimensions are out-of-bounds (i.e. the default
+    value is used).
+
+    A `vector.transfer_read` can be lowered to a simple load if all dimensions
+    are specified to be within bounds and no `mask` was specified.
 
     This operation is called 'read' by opposition to 'load' because the
     super-vector granularity is generally not representable with a single
@@ -1607,7 +1610,7 @@ def Vector_TransferWriteOp :
                    Variadic<Index>:$indices,
                    AffineMapAttr:$permutation_map,
                    Optional<VectorOf<[I1]>>:$mask,
-                   OptionalAttr<BoolArrayAttr>:$in_bounds)>,
+                   BoolArrayAttr:$in_bounds)>,
     Results<(outs Optional<AnyRankedTensor>:$result)> {
 
   let summary = "The vector.transfer_write op writes a supervector to memory.";
@@ -1643,15 +1646,19 @@ def Vector_TransferWriteOp :
     any permutation. Elements whose corresponding mask element is `0` are
     masked out.
 
-    An optional boolean array attribute `in_bounds` specifies for every vector
-    dimension if the transfer is guaranteed to be within the source bounds. If
-    specified, the `in_bounds` array length has to be equal to the vector rank.
-    If set to "false", accesses (including the starting point) may run
+    For every vector dimension, the boolean array attribute `in_bounds`
+    specifies if the transfer is guaranteed to be within the source bounds. If
+    set to "false", accesses (including the starting point) may run
     out-of-bounds along the respective vector dimension as the index increases.
-    In absence of the attribute, accesses along all vector dimensions may run
-    out-of-bounds. A `vector.transfer_write` can be lowered to a simple store if
-    all dimensions are specified to be within bounds and no `mask` was
-    specified. Note that non-vector dimensions *must* always be in-bounds.
+    Non-vector and broadcast dimensions *must* always be in-bounds. The
+    `in_bounds` array length has to be equal to the vector rank. This attribute
+    has a default value: `false` (i.e. "out-of-bounds"). When skipped in the
+    textual IR, the default value is assumed. Similarly, the OP printer will
+    omit this attribute when all dimensions are out-of-bounds (i.e. the default
+    value is used).
+
+     A `vector.transfer_write` can be lowered to a simple store if all
+     dimensions are specified to be within bounds and no `mask` was specified.
 
     This operation is called 'write' by opposition to 'store' because the
     super-vector granularity is generally not representable with a single

mlir/include/mlir/IR/AffineMap.h

@@ -146,6 +146,14 @@ class AffineMap {
   /// affine map (d0, ..., dn) -> (dp, ..., dn) on the most minor dimensions.
   bool isMinorIdentity() const;
 
+  /// Returns the list of broadcast dimensions (i.e. dims indicated by value 0
+  /// in the result).
+  /// Ex:
+  ///  * (d0, d1, d2) -> (0, d1) gives [0]
+  ///  * (d0, d1, d2) -> (d2, d1) gives []
+  ///  * (d0, d1, d2, d4) -> (d0, 0, d1, 0) gives [1, 3]
+  SmallVector<unsigned> getBroadcastDims() const;
+
   /// Returns true if this affine map is a minor identity up to broadcasted
   /// dimensions which are indicated by value 0 in the result. If
   /// `broadcastedDims` is not null, it will be populated with the indices of

mlir/include/mlir/Interfaces/VectorInterfaces.td

@@ -98,7 +98,7 @@ def VectorTransferOpInterface : OpInterface<"VectorTransferOpInterface"> {
         dimension whether it is in-bounds or not. (Broadcast dimensions are
         always in-bounds).
       }],
-      /*retTy=*/"::std::optional<::mlir::ArrayAttr>",
+      /*retTy=*/"::mlir::ArrayAttr",
       /*methodName=*/"getInBounds",
       /*args=*/(ins)
     >,
@@ -240,9 +240,7 @@ def VectorTransferOpInterface : OpInterface<"VectorTransferOpInterface"> {
     bool isDimInBounds(unsigned dim) {
       if ($_op.isBroadcastDim(dim))
         return true;
-      if (!$_op.getInBounds())
-        return false;
-      auto inBounds = ::llvm::cast<::mlir::ArrayAttr>(*$_op.getInBounds());
+      auto inBounds = $_op.getInBounds();
       return ::llvm::cast<::mlir::BoolAttr>(inBounds[dim]).getValue();
     }
 

mlir/lib/Dialect/Affine/Transforms/SuperVectorize.cpp

@@ -1223,8 +1223,19 @@ static Operation *vectorizeAffineLoad(AffineLoadOp loadOp,
   LLVM_DEBUG(dbgs() << "\n[early-vect]+++++ permutationMap: ");
   LLVM_DEBUG(permutationMap.print(dbgs()));
 
+  // Make sure that the in_bounds attribute corresponding to a broadcast dim
+  // is set to `true` - that's required by the xfer Op.
+  // FIXME: We're not veryfying whether the corresponding access is in bounds.
+  // TODO: Use masking instead.
+  SmallVector<unsigned> broadcastedDims = permutationMap.getBroadcastDims();
+  SmallVector<bool> inBounds(vectorType.getRank(), false);
+
+  for (auto idx : broadcastedDims)
+    inBounds[idx] = true;
+
   auto transfer = state.builder.create<vector::TransferReadOp>(
-      loadOp.getLoc(), vectorType, loadOp.getMemRef(), indices, permutationMap);
+      loadOp.getLoc(), vectorType, loadOp.getMemRef(), indices, permutationMap,
+      inBounds);
 
   // Register replacement for future uses in the scope.
   state.registerOpVectorReplacement(loadOp, transfer);

mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp

@@ -1343,8 +1343,17 @@ vectorizeAsLinalgGeneric(RewriterBase &rewriter, VectorizationState &state,
 
     SmallVector<Value> indices(linalgOp.getShape(opOperand).size(), zero);
 
+    // Make sure that the in_bounds attribute corresponding to a broadcast dim
+    // is `true`
+    SmallVector<unsigned> broadcastedDims = readMap.getBroadcastDims();
+    SmallVector<bool> inBounds(readType.getRank(), false);
+
+    for (auto idx : broadcastedDims)
+      inBounds[idx] = true;
+
     Operation *read = rewriter.create<vector::TransferReadOp>(
-        loc, readType, opOperand->get(), indices, readMap);
+        loc, readType, opOperand->get(), indices, readMap,
+        ArrayRef<bool>(inBounds));
     read = state.maskOperation(rewriter, read, linalgOp, maskingMap);
     Value readValue = read->getResult(0);
 
@@ -2681,11 +2690,12 @@ LogicalResult LinalgCopyVTRForwardingPattern::matchAndRewrite(
   // The `masked` attribute is only valid on this padded buffer.
   // When forwarding to vector.transfer_read, the attribute must be reset
   // conservatively.
+  auto vectorType = xferOp.getVectorType();
   Value res = rewriter.create<vector::TransferReadOp>(
-      xferOp.getLoc(), xferOp.getVectorType(), in, xferOp.getIndices(),
+      xferOp.getLoc(), vectorType, in, xferOp.getIndices(),
       xferOp.getPermutationMapAttr(), xferOp.getPadding(), xferOp.getMask(),
-      // in_bounds is explicitly reset
-      /*inBoundsAttr=*/ArrayAttr());
+      rewriter.getBoolArrayAttr(
+          SmallVector<bool>(vectorType.getRank(), false)));
 
   if (maybeFillOp)
     rewriter.eraseOp(maybeFillOp);
@@ -2739,11 +2749,12 @@ LogicalResult LinalgCopyVTWForwardingPattern::matchAndRewrite(
   // The `masked` attribute is only valid on this padded buffer.
   // When forwarding to vector.transfer_write, the attribute must be reset
   // conservatively.
+  auto vector = xferOp.getVector();
   rewriter.create<vector::TransferWriteOp>(
-      xferOp.getLoc(), xferOp.getVector(), out, xferOp.getIndices(),
+      xferOp.getLoc(), vector, out, xferOp.getIndices(),
       xferOp.getPermutationMapAttr(), xferOp.getMask(),
-      // in_bounds is explicitly reset
-      /*inBoundsAttr=*/ArrayAttr());
+      rewriter.getBoolArrayAttr(
+          SmallVector<bool>(vector.getType().getRank(), false)));
 
   rewriter.eraseOp(copyOp);
   rewriter.eraseOp(xferOp);

mlir/lib/Dialect/Vector/IR/VectorOps.cpp

@@ -3818,7 +3818,8 @@ void TransferReadOp::build(OpBuilder &builder, OperationState &result,
   auto permutationMapAttr = AffineMapAttr::get(permutationMap);
   auto inBoundsAttr = (inBounds && !inBounds.value().empty())
                           ? builder.getBoolArrayAttr(inBounds.value())
-                          : ArrayAttr();
+                          : builder.getBoolArrayAttr(
+                                SmallVector<bool>(vectorType.getRank(), false));
   build(builder, result, vectorType, source, indices, permutationMapAttr,
         inBoundsAttr);
 }
@@ -3833,7 +3834,8 @@ void TransferReadOp::build(OpBuilder &builder, OperationState &result,
   auto permutationMapAttr = AffineMapAttr::get(permutationMap);
   auto inBoundsAttr = (inBounds && !inBounds.value().empty())
                           ? builder.getBoolArrayAttr(inBounds.value())
-                          : ArrayAttr();
+                          : builder.getBoolArrayAttr(
+                                SmallVector<bool>(vectorType.getRank(), false));
   build(builder, result, vectorType, source, indices, permutationMapAttr,
         padding,
         /*mask=*/Value(), inBoundsAttr);
@@ -3951,17 +3953,15 @@ verifyTransferOp(VectorTransferOpInterface op, ShapedType shapedType,
            << inferredMaskType << ") and mask operand type (" << maskType
            << ") don't match";
 
-  if (inBounds) {
-    if (permutationMap.getNumResults() != static_cast<int64_t>(inBounds.size()))
-      return op->emitOpError("expects the optional in_bounds attr of same rank "
-                             "as permutation_map results: ")
-             << AffineMapAttr::get(permutationMap)
-             << " vs inBounds of size: " << inBounds.size();
-    for (unsigned int i = 0; i < permutationMap.getNumResults(); ++i)
-      if (isa<AffineConstantExpr>(permutationMap.getResult(i)) &&
-          !llvm::cast<BoolAttr>(inBounds.getValue()[i]).getValue())
-        return op->emitOpError("requires broadcast dimensions to be in-bounds");
-  }
+  if (permutationMap.getNumResults() != static_cast<int64_t>(inBounds.size()))
+    return op->emitOpError("expects the in_bounds attr of same rank "
+                           "as permutation_map results: ")
+           << AffineMapAttr::get(permutationMap)
+           << " vs inBounds of size: " << inBounds.size();
+  for (unsigned int i = 0, e = permutationMap.getNumResults(); i < e; ++i)
+    if (isa<AffineConstantExpr>(permutationMap.getResult(i)) &&
+        !llvm::cast<BoolAttr>(inBounds.getValue()[i]).getValue())
+      return op->emitOpError("requires broadcast dimensions to be in-bounds");
 
   return success();
 }
@@ -4037,6 +4037,13 @@ ParseResult TransferReadOp::parse(OpAsmParser &parser, OperationState &result) {
   } else {
     permMap = llvm::cast<AffineMapAttr>(permMapAttr).getValue();
   }
+  auto inBoundsAttrName = TransferReadOp::getInBoundsAttrName(result.name);
+  Attribute inBoundsAttr = result.attributes.get(inBoundsAttrName);
+  if (!inBoundsAttr) {
+    result.addAttribute(inBoundsAttrName,
+                        builder.getBoolArrayAttr(
+                            SmallVector<bool>(permMap.getNumResults(), false)));
+  }
   if (parser.resolveOperand(sourceInfo, shapedType, result.operands) ||
       parser.resolveOperands(indexInfo, indexType, result.operands) ||
       parser.resolveOperand(paddingInfo, shapedType.getElementType(),
@@ -4081,8 +4088,7 @@ LogicalResult TransferReadOp::verify() {
 
   if (failed(verifyTransferOp(cast<VectorTransferOpInterface>(getOperation()),
                               shapedType, vectorType, maskType,
-                              inferredMaskType, permutationMap,
-                              getInBounds() ? *getInBounds() : ArrayAttr())))
+                              inferredMaskType, permutationMap, getInBounds())))
     return failure();
 
   if (auto sourceVectorElementType =
@@ -4355,9 +4361,11 @@ void TransferWriteOp::build(OpBuilder &builder, OperationState &result,
                             AffineMap permutationMap,
                             std::optional<ArrayRef<bool>> inBounds) {
   auto permutationMapAttr = AffineMapAttr::get(permutationMap);
-  auto inBoundsAttr = (inBounds && !inBounds.value().empty())
-                          ? builder.getBoolArrayAttr(inBounds.value())
-                          : ArrayAttr();
+  auto inBoundsAttr =
+      (inBounds && !inBounds.value().empty())
+          ? builder.getBoolArrayAttr(inBounds.value())
+          : builder.getBoolArrayAttr(SmallVector<bool>(
+                llvm::cast<VectorType>(vector.getType()).getRank(), false));
   build(builder, result, vector, dest, indices, permutationMapAttr,
         /*mask=*/Value(), inBoundsAttr);
 }
@@ -4410,6 +4418,13 @@ ParseResult TransferWriteOp::parse(OpAsmParser &parser,
   } else {
     permMap = llvm::cast<AffineMapAttr>(permMapAttr).getValue();
   }
+  auto inBoundsAttrName = TransferWriteOp::getInBoundsAttrName(result.name);
+  Attribute inBoundsAttr = result.attributes.get(inBoundsAttrName);
+  if (!inBoundsAttr) {
+    result.addAttribute(inBoundsAttrName,
+                        builder.getBoolArrayAttr(
+                            SmallVector<bool>(permMap.getNumResults(), false)));
+  }
   if (parser.resolveOperand(vectorInfo, vectorType, result.operands) ||
       parser.resolveOperand(sourceInfo, shapedType, result.operands) ||
       parser.resolveOperands(indexInfo, indexType, result.operands))
@@ -4463,8 +4478,7 @@ LogicalResult TransferWriteOp::verify() {
 
   if (failed(verifyTransferOp(cast<VectorTransferOpInterface>(getOperation()),
                               shapedType, vectorType, maskType,
-                              inferredMaskType, permutationMap,
-                              getInBounds() ? *getInBounds() : ArrayAttr())))
+                              inferredMaskType, permutationMap, getInBounds())))
     return failure();
 
   return verifyPermutationMap(permutationMap,

mlir/lib/Dialect/Vector/Transforms/LowerVectorMask.cpp

@@ -224,7 +224,7 @@ struct MaskedTransferReadOpPattern
     rewriter.replaceOpWithNewOp<TransferReadOp>(
         maskingOp.getOperation(), readOp.getVectorType(), readOp.getSource(),
         readOp.getIndices(), readOp.getPermutationMap(), readOp.getPadding(),
-        maskingOp.getMask(), readOp.getInBounds().value_or(ArrayAttr()));
+        maskingOp.getMask(), readOp.getInBounds());
     return success();
   }
 };
@@ -246,7 +246,7 @@ struct MaskedTransferWriteOpPattern
     rewriter.replaceOpWithNewOp<TransferWriteOp>(
         maskingOp.getOperation(), resultType, writeOp.getVector(),
         writeOp.getSource(), writeOp.getIndices(), writeOp.getPermutationMap(),
-        maskingOp.getMask(), writeOp.getInBounds().value_or(ArrayAttr()));
+        maskingOp.getMask(), writeOp.getInBounds());
     return success();
   }
 };

mlir/lib/Dialect/Vector/Transforms/LowerVectorTransfer.cpp

@@ -133,9 +133,7 @@ struct TransferReadPermutationLowering
 
     // Transpose in_bounds attribute.
     ArrayAttr newInBoundsAttr =
-        op.getInBounds() ? inverseTransposeInBoundsAttr(
-                               rewriter, op.getInBounds().value(), permutation)
-                         : ArrayAttr();
+        inverseTransposeInBoundsAttr(rewriter, op.getInBounds(), permutation);
 
     // Generate new transfer_read operation.
     VectorType newReadType = VectorType::get(
@@ -208,9 +206,7 @@ struct TransferWritePermutationLowering
 
     // Transpose in_bounds attribute.
     ArrayAttr newInBoundsAttr =
-        op.getInBounds() ? inverseTransposeInBoundsAttr(
-                               rewriter, op.getInBounds().value(), permutation)
-                         : ArrayAttr();
+        inverseTransposeInBoundsAttr(rewriter, op.getInBounds(), permutation);
 
     // Generate new transfer_write operation.
     Value newVec = rewriter.create<vector::TransposeOp>(

mlir/lib/IR/AffineMap.cpp

@@ -158,6 +158,19 @@ bool AffineMap::isMinorIdentity() const {
              getMinorIdentityMap(getNumDims(), getNumResults(), getContext());
 }
 
+SmallVector<unsigned> AffineMap::getBroadcastDims() const {
+  SmallVector<unsigned> broadcastedDims;
+  for (const auto &[resIdx, expr] : llvm::enumerate(getResults())) {
+    if (auto constExpr = dyn_cast<AffineConstantExpr>(expr)) {
+      if (constExpr.getValue() != 0)
+        continue;
+      broadcastedDims.push_back(resIdx);
+    }
+  }
+
+  return broadcastedDims;
+}
+
 /// Returns true if this affine map is a minor identity up to broadcasted
 /// dimensions which are indicated by value 0 in the result.
 bool AffineMap::isMinorIdentityWithBroadcasting(

mlir/test/Conversion/VectorToSCF/vector-to-scf.mlir

@@ -133,7 +133,7 @@ func.func @materialize_read(%M: index, %N: index, %O: index, %P: index) {
     affine.for %i1 = 0 to %N {
       affine.for %i2 = 0 to %O {
         affine.for %i3 = 0 to %P step 5 {
-          %f = vector.transfer_read %A[%i0, %i1, %i2, %i3], %f0 {permutation_map = affine_map<(d0, d1, d2, d3) -> (d3, 0, d0)>} : memref<?x?x?x?xf32>, vector<5x4x3xf32>
+          %f = vector.transfer_read %A[%i0, %i1, %i2, %i3], %f0 {in_bounds = [false, true, false], permutation_map = affine_map<(d0, d1, d2, d3) -> (d3, 0, d0)>} : memref<?x?x?x?xf32>, vector<5x4x3xf32>
           // Add a dummy use to prevent dead code elimination from removing
           // transfer read ops.
           "dummy_use"(%f) : (vector<5x4x3xf32>) -> ()
@@ -507,7 +507,7 @@ func.func @transfer_read_with_tensor(%arg: tensor<f32>) -> vector<1xf32> {
     // CHECK-NEXT: %[[RESULT:.*]] = vector.broadcast %[[EXTRACTED]] : f32 to vector<1xf32>
     // CHECK-NEXT: return %[[RESULT]] : vector<1xf32>
     %f0 = arith.constant 0.0 : f32
-    %0 = vector.transfer_read %arg[], %f0 {permutation_map = affine_map<()->(0)>} :
+    %0 = vector.transfer_read %arg[], %f0 {in_bounds = [true], permutation_map = affine_map<()->(0)>} :
       tensor<f32>, vector<1xf32>
     return %0: vector<1xf32>
 }
@@ -746,7 +746,7 @@ func.func @cannot_lower_transfer_read_with_leading_scalable(%arg0: memref<?x4xf3
 func.func @does_not_crash_on_unpack_one_dim(%subview:  memref<1x1x1x1xi32>, %mask: vector<1x1xi1>) -> vector<1x1x1x1xi32> {
   %c0 = arith.constant 0 : index
   %c0_i32 = arith.constant 0 : i32
-  %3 = vector.transfer_read %subview[%c0, %c0, %c0, %c0], %c0_i32, %mask {permutation_map = #map1}
+  %3 = vector.transfer_read %subview[%c0, %c0, %c0, %c0], %c0_i32, %mask {in_bounds = [false, true, true, false], permutation_map = #map1}
           : memref<1x1x1x1xi32>, vector<1x1x1x1xi32>
   return %3 : vector<1x1x1x1xi32>
 }