[MLIR] Legalize certain vector.transfer_read ops of scalable vectors (#143146)

This patch adds a transform of `transfer_read` operation to change the
vector type to one that can be mapped to an LLVM type. This is done by
collapsing trailing dimensions so we obtain a vector type with a single
trailing scalable dimension.

Author: momchil-velikov

mlir/lib/Dialect/ArmSVE/Transforms/LegalizeVectorStorage.cpp

@@ -298,16 +298,156 @@ struct LegalizeSVEMaskLoadConversion : public OpRewritePattern<memref::LoadOp> {
   }
 };
 
+/// Transforms a `transfer_read` operation so it reads vector of a type that
+/// can be mapped to an LLVM type ("LLVM-legal" type). This is done by
+/// collapsing trailing dimensions so we obtain a vector type with a single
+/// scalable dimension in the rightmost position.
+///
+/// Example:
+/// ```
+/// %v = vector.transfer_read %M[%i, %j, %c0, %c0], %c0_i8
+///   {in_bounds = [false, true, true, true]}
+///   : memref<?x?x2x8xi8>, vector<2x[4]x2x8xi8>
+/// ```
+/// is rewritten to
+/// ```
+/// %collapse_shape = memref.collapse_shape %M [[0], [1, 2, 3]]
+///   : memref<?x?x2x8xi8> into memref<?x?xi8>
+/// %0 = vector.transfer_read  %collapse_shape[%i, %j], %c0_i8
+///   {in_bounds = [false, true]}
+///   : memref<?x?xi8>, vector<2x[64]xi8>
+/// %1 = vector.shape_cast %0 : vector<2x[64]xi8> to vector<2x[4]x2x8xi8>
+/// ```
+struct LegalizeTransferRead : public OpRewritePattern<vector::TransferReadOp> {
+  using OpRewritePattern::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(vector::TransferReadOp readOp,
+                                PatternRewriter &rewriter) const override {
+
+    // Do not try to transform masked reads. For example, if we have a transfer
+    // to a `vector<[4]x4xi8>` we could have a mask like
+    //    1 1 1 0
+    //    1 1 1 0
+    //    1 1 1 0
+    //    0 0 0 0
+    // Flattening this mask would look like
+    //    1 1 1 0 1 1 1 0 1 1 1 0 0 0 0 0
+    // and we have not yet figured out an efficient way to build such a mask,
+    // neither from the mask operand, nor from the original `vector.create_mask`
+    // operation (if visible at all).
+    if (readOp.isMasked() || readOp.getMask())
+      return rewriter.notifyMatchFailure(readOp,
+                                         "masked transfers not-supported");
+
+    // General permutation maps are not supported. The issue is with transpose,
+    // broadcast, and other forms of non-identify mapping in the minor
+    // dimensions which is impossible to represent after collapsing (at least
+    // because the resulting "collapsed" maps would have smaller number of
+    // dimension indices).
+    // TODO: We have not had yet the need for it, but some forms of permutation
+    // maps with identity in the minor dimensions voukld be supported, for
+    // example `(i, j, k, p) -> (j, i, k, p)` where we need to collapse only `k`
+    // and `p`.
+    if (!readOp.getPermutationMap().isMinorIdentity())
+      return rewriter.notifyMatchFailure(readOp, "non-identity permutation");
+
+    // We handle transfers of vectors with rank >= 2 and a single scalable
+    // dimension. This transformation aims to transform an LLVM-illegal type
+    // into an LLVM-legal type and one dimensional vectors are already
+    // LLVM-legal, even if scalable. A value of a vector type with more than one
+    // scalable dimension is impossible to represent using a vector type with no
+    // scalable dimensions or a single one. For example a `vector<[4]x[4]xi8>`
+    // would have `4 * 4 * vscale * vscale` elements and this quantity is
+    // impossible to represent as `N` or `N * vscale` (where `N` is a constant).
+    VectorType origVT = readOp.getVectorType();
+    ArrayRef<bool> origScalableDims = origVT.getScalableDims();
+    const int64_t origVRank = origVT.getRank();
+    if (origVRank < 2 || origVT.getNumScalableDims() != 1)
+      return rewriter.notifyMatchFailure(readOp, "wrong dimensions");
+
+    // Number of trailing dimensions to collapse, including the scalable
+    // dimension.  Nothing to do if the single scalable dimension is already the
+    // last one.
+    const int64_t numCollapseDims = std::distance(
+        llvm::find(origScalableDims, true), origScalableDims.end());
+    if (numCollapseDims < 2)
+      return rewriter.notifyMatchFailure(readOp,
+                                         "scalable dimension is trailing");
+
+    // We want a simple memref (not a tensor) with contiguous elements for at
+    // least all the trailing dimensions up to and including the scalable one.
+    auto memTy = dyn_cast<MemRefType>(readOp.getBase().getType());
+    if (!(memTy && memTy.areTrailingDimsContiguous(numCollapseDims)))
+      return rewriter.notifyMatchFailure(
+          readOp, "non-contiguous memref dimensions to collapse");
+
+    // The dimensions to collapse (excluding the scalable one) of the vector and
+    // the memref must match. A dynamic memref dimension is considered
+    // non-matching. The transfers from the dimensions to collapse must be
+    // in-bounds (it follows the corresponding indices would be zero). This
+    // guarantees that the operation transfers a contiguous block
+    // and no padding is necessary.
+    if (!llvm::equal(memTy.getShape().take_back(numCollapseDims - 1),
+                     origVT.getShape().take_back(numCollapseDims - 1)))
+      return rewriter.notifyMatchFailure(
+          readOp, "memref and vector dimensions do not match");
+
+    SmallVector<bool> origInBounds = readOp.getInBoundsValues();
+    if (!llvm::all_of(
+            ArrayRef<bool>(origInBounds).take_back(numCollapseDims - 1),
+            [](bool v) { return v; }))
+      return rewriter.notifyMatchFailure(
+          readOp, "out-of-bounds transfer from a dimension to collapse");
+
+    // Collapse the trailing dimensions of the memref.
+    SmallVector<ReassociationIndices> reassoc;
+    for (int64_t i = 0; i < memTy.getRank() - numCollapseDims + 1; ++i)
+      reassoc.push_back({i});
+    for (int64_t i = memTy.getRank() - numCollapseDims + 1; i < memTy.getRank();
+         ++i)
+      reassoc.back().push_back(i);
+    if (!memref::CollapseShapeOp::isGuaranteedCollapsible(memTy, reassoc))
+      return failure();
+    Value collapsedMem = rewriter.create<memref::CollapseShapeOp>(
+        readOp.getLoc(), readOp.getBase(), reassoc);
+
+    // Get a vector type with collapsed trailing dimensions.
+    SmallVector<int64_t> shape(origVT.getShape());
+    for (int64_t i = origVRank - numCollapseDims + 1; i < origVRank; ++i)
+      shape[origVRank - numCollapseDims] *= shape[i];
+    shape.pop_back_n(numCollapseDims - 1);
+    auto collapsedVT =
+        VectorType::get(shape, origVT.getElementType(),
+                        origScalableDims.drop_back(numCollapseDims - 1));
+
+    // Drop the extra (zero) indices.
+    auto indices = readOp.getIndices().drop_back(numCollapseDims - 1);
+
+    // Create the new `transfer_read`.
+    auto newReadOp = rewriter.create<vector::TransferReadOp>(
+        readOp.getLoc(), collapsedVT, collapsedMem, indices,
+        ArrayRef<bool>(origInBounds).drop_back(numCollapseDims - 1));
+
+    // Cast back to the original vector type.
+    auto toOrigShape = rewriter.create<vector::ShapeCastOp>(readOp.getLoc(),
+                                                            origVT, newReadOp);
+
+    rewriter.replaceOp(readOp, toOrigShape);
+    return success();
+  }
+};
+
 } // namespace
 
 void mlir::arm_sve::populateLegalizeVectorStoragePatterns(
     RewritePatternSet &patterns) {
-  patterns.add<RelaxScalableVectorAllocaAlignment,
-               LegalizeSVEMaskAllocation<memref::AllocaOp>,
-               LegalizeSVEMaskAllocation<memref::AllocOp>,
-               LegalizeSVEMaskTypeCastConversion,
-               LegalizeSVEMaskStoreConversion, LegalizeSVEMaskLoadConversion>(
-      patterns.getContext());
+  patterns
+      .add<RelaxScalableVectorAllocaAlignment,
+           LegalizeSVEMaskAllocation<memref::AllocaOp>,
+           LegalizeSVEMaskAllocation<memref::AllocOp>,
+           LegalizeSVEMaskTypeCastConversion, LegalizeSVEMaskStoreConversion,
+           LegalizeSVEMaskLoadConversion, LegalizeTransferRead>(
+          patterns.getContext());
 }
 
 namespace {