[mlir][ArmSME] Add move vector to tile slice op and lowerings

[c-rhodes]

Summary

This adds a 'move_vector_to_tile_slice' op to the ArmSME dialect that
moves a 1-D scalable vector to a slice of a 2-D tile at a given index.

This is lowered to the 'llvm.aarch64.sme.write.horiz' intrinsic that
maps to the MOVA (vector to tile, single) SME instruction [1] when
lowering to LLVM. Like the SME load and store instructions this operates
on ZA tile slices, which are 1D vectors of horizontally or vertically
contiguous elements within a ZA tile.

This patch extends the lowering of 'arith.constant' to SME to support
non-zero constants using this new op. This requires materializing a
loop that broadcasts the constant to each tile slice with the
'vector_to_tile_slice' op. Unlike load and store, this is done during
conversion from Vector to ArmSME, rather than ArmSME to SCF. The latter
would require a higher-level custom op in the ArmSME dialect like
'tile_load' and 'tile_store' and this isn't necessary. We may also
remove the load and store ops in the future in favour of lowering
straight from Vector, at which point this would converge.

Currently only horizontal tile slices are supported. A future patch will
extend this mechanism to support 'vector.broadcast'.

Depends on D156980 D157004

[1] https://developer.arm.com/documentation/ddi0602

[dcaballe]

LGTM in general, thanks!

mlir/include/mlir/Dialect/ArmSME/IR/ArmSME.td

424

-> element type?

mlir/lib/Dialect/ArmSME/Transforms/LegalizeForLLVMExport.cpp

340–345

We are usually creating masks with vector.create_mask and vector.constant_mask but maybe it's already too late to introduce them here. I would expect them to have been lowered in an earlier stage. Just bringing this up in case you thought about it. (I'm not asking for changes)

[awarzynski]

A few minor points/questions. LG otherwise

mlir/include/mlir/Dialect/ArmSME/IR/ArmSME.td

417

Most Ops' names convey what the corresponding "action"/"function" is. Wondering whether this shouldn't be MoveVectorToTileSliceOp instead. Naming is hard.

mlir/lib/Conversion/VectorToArmSME/VectorToArmSME.cpp

106

[nit] Just to make the split into 2 cases more visible.

112

[nit] Just to make the split into 2 cases more visible.

137

IIUC, the following block will only create the loop structure. The following operation is not created here:

loads each ZA tile slice from memory.

That's created further down. Also, rather than "loading from memory", this is creating "move from vector to a slice", right?

mlir/test/Dialect/ArmSME/roundtrip.mlir

586

Could you add some invalid cases in "invalid.mlir"?

mlir/test/Dialect/ArmSME/vector-ops-to-sme.mlir

251 ↗

(On Diff #546856)

It would be good to check at least one more element type.

mlir/test/Integration/Dialect/Vector/CPU/ArmSME/tile_fill.mlir

55

[nit] I would use some other value - 1 is super common and can be easily missed. Here it would be nice to emphasise that it could be _anything_.

[benmxwl-arm]

mlir/test/Integration/Dialect/Vector/CPU/ArmSME/tile_fill.mlir

75

It's landed :)

[c-rhodes]

mlir/include/mlir/Dialect/ArmSME/IR/ArmSME.td

417

Most Ops' names convey what the corresponding "action"/"function" is. Wondering whether this shouldn't be MoveVectorToTileSliceOp instead. Naming is hard.

are you suggesting to rename it both internally and externally (i.e. move_vector_to_tile_slice) or just the former?

424

-> element type?

For a 2-d scalable vector such as vector<[4]x[4]xi32> is the element type vector<[4]xi32>? That's what I want to express here but I do struggle with these descriptions. FWIW getElementType() returns the scalar i32. Perhaps The 1-D vector type must match the vector type of the inner dimension of the 2-D vector type?

mlir/test/Integration/Dialect/Vector/CPU/ArmSME/tile_fill.mlir

75

It's landed :)

Thanks for heads up I'll update this

[c-rhodes]

• Address most comments.
• Add type constraint to VectorToTileSliceOp that verifies 1-D vector type matches element type of 2-D vector type (tile).
• Rename VectorToTileSliceOp::getVectorType() -> VectorToTileSliceOp::getTileType() to prevent confusion with vector operand.

[c-rhodes]

mlir/lib/Conversion/VectorToArmSME/VectorToArmSME.cpp

137

IIUC, the following block will only create the loop structure. The following operation is not created here:

loads each ZA tile slice from memory.

That's created further down. Also, rather than "loading from memory", this is creating "move from vector to a slice", right?

mlir/lib/Dialect/ArmSME/Transforms/LegalizeForLLVMExport.cpp

340–345

We are usually creating masks with vector.create_mask and vector.constant_mask but maybe it's already too late to introduce them here. I would expect them to have been lowered in an earlier stage. Just bringing this up in case you thought about it. (I'm not asking for changes)

I hadn't thought about that, thanks for mentioning. I can't see why it wouldn't work since a vector op (splat) is already added here, but I'm not really sure it simplifies this.

mlir/test/Dialect/ArmSME/roundtrip.mlir

586

Could you add some invalid cases in "invalid.mlir"?

Done, was also missing a type constraint that verifies 1-d vector type matches inner vector type of 2-d vector type.

[awarzynski]

mlir/include/mlir/Dialect/ArmSME/IR/ArmSME.td

417

I was thinking both.

424

I also though that you meant the element type here, as in e.g. i32 or f32.

I think that we need to be more explicit here:

The type of the 1-d scalable vector to be moved must match the type of the tile slice (note that 1 slice is effectively 1 row in a virtual tile).

WDYT?

[c-rhodes]

Update descriptions for op and type constraint.

[c-rhodes]

mlir/include/mlir/Dialect/ArmSME/IR/ArmSME.td

417

I was thinking both.

Still considering, it's more descriptive but also it's quite a long name

424

I also though that you meant the element type here, as in e.g. i32 or f32.

I think that we need to be more explicit here:

The type of the 1-d scalable vector to be moved must match the type of the tile slice (note that 1 slice is effectively 1 row in a virtual tile).

WDYT?

I've updated the description along these lines, hope it makes more sense now, appreciate the input!

[c-rhodes]

Rename vector_to_tile_slice to move_vector_to_tile_slice

[dcaballe]

mlir/include/mlir/Dialect/ArmSME/IR/ArmSME.td

424

I would call that inner 1-D sub-vector?

430

1-d -> 1-D for consistency?

mlir/lib/Dialect/ArmSME/Transforms/LegalizeForLLVMExport.cpp

340–345

Well, if this is expected to run before some of these "higher" level vector ops are lowered, I would make sure we generate those. We have patterns that are looking specifically for those and are able to understand if they are all-true/all-false masks, etc. We don't have mask patterns looking at constant ops.

[awarzynski]

LGTM, thanks!

Following on from what Diego suggest re vector.mask, we probably should look into using those instead of vector.splat. At some point soon, not in this patch :)

[c-rhodes]

mlir/include/mlir/Dialect/ArmSME/IR/ArmSME.td

424

I would call that inner 1-D sub-vector?

Hm, it's not an inner 1-D sub-vector until it's moved, I think that would make sense for the inverse of this operation (extract) that does tile slice to vector.

430

1-d -> 1-D for consistency?

Ah good spot will fix this before landing, cheers

mlir/lib/Dialect/ArmSME/Transforms/LegalizeForLLVMExport.cpp

340–345

Well, if this is expected to run before some of these "higher" level vector ops are lowered, I would make sure we generate those. We have patterns that are looking specifically for those and are able to understand if they are all-true/all-false masks, etc. We don't have mask patterns looking at constant ops.

Sorry I'm not sure I follow, this is low-level and emitting the canonical form for an all active mask (a constant?), it's not clear to me what the benefit of such patterns are in this case?

I should mention there has been no consideration of masking so far in the ops introduced for our first target for SME of lowering linalg.fill, now that there is a basic path established (D158619) I've starting looking into masking.

[nicolasvasilache]

mlir/test/Dialect/ArmSME/arith-ops-to-sme.mlir

107

side note, can you just return the value instead of using this fake op or is there something more fundamental that does not let us return a scalable vector here ?

[c-rhodes]

mlir/test/Dialect/ArmSME/arith-ops-to-sme.mlir

107

side note, can you just return the value instead of using this fake op or is there something more fundamental that does not let us return a scalable vector here ?

For the purposes of this test the scalable vector could be returned, but generally we can't support passing or returning 2-d scalable vectors to/from functions since these types can't be lowered to LLVM and even if they could it's not defined by the ABI. For this reason I opted for the fake use op so as to not set a precedent that this is something that can be done. I believe there are some earlier ArmSME tests where 2-d scalable vector are returned however, we should probably update them for consistency.

mlir/include/mlir/Dialect/ArmSME/IR/ArmSME.td

def StoreTileSliceOp : ArmSME_Op<"store_tile_slice"> {
   }];
 
   let assemblyFormat = [{
     $tile `,` $tile_slice_index `,` $base `[` $indices `]`
       attr-dict `:` type($base) `,` type($tile)
   }];
 }
 
+def MoveVectorToTileSliceOp : ArmSME_Op<"move_vector_to_tile_slice", [
+    AllTypesMatch<["tile", "result"]>,
+    TypesMatchWith<
+      "type of 'vector' matches type of 'tile' slice",
+      "tile", "vector",
+      "VectorType::get("
+        "::llvm::cast<mlir::VectorType>($_self).getShape().drop_front(),"
+        "::llvm::cast<mlir::VectorType>($_self).getElementType(),"
+        "/*scalableDims=*/{true})">,
+]> {
+  let summary = "Move 1-D scalable vector to slice of 2-D tile";
+  let description = [{
+    The vector to tile slice operation moves a 1-D scalable vector to a slice
+    of a 2-D scalable vector tile at the given index. The type of the 1-D
+    scalable vector to be moved must match the type of the tile slice. A tile
+    slice is a 1-D vector of horizontally or vertically contiguous elements
+    within a ZA tile. Horizontal tile slices are currently assumed when
+    lowering to intrinsics. The updated tile is returned as the result.
+
+    Example 1: Move a vector<[16]xi8> into tile at given index.
+    ```mlir
+    %tile_update = arm_sme.move_vector_to_tile_slice %vector, %tile, %tile_slice_index : vector<[16]xi8> into vector<[16]x[16]xi8>
+    ```
+
+    Example 2: Move a vector<[2]xf64> into tile at given index.
+    ```mlir
+    %tile_update = arm_sme.move_vector_to_tile_slice %vector, %tile, %tile_slice_index : vector<[2]xf64> into vector<[2]x[2]xf64>
+    ```
+  }];
+  let arguments = (ins
+      SVEVector:$vector, SMETile:$tile, Index:$tile_slice_index);
+  let results = (outs SMETile:$result);
+
+  let extraClassDeclaration = [{
+    VectorType getTileType() {
+      return ::llvm::cast<VectorType>(getTile().getType());
+    }
+  }];
+
+  let assemblyFormat = [{
+    $vector `,` $tile `,` $tile_slice_index
+      attr-dict `:` type($vector) `into` type($result)
+  }];
+}
+
 //===----------------------------------------------------------------------===//
 // ArmSME Intrinsic op definitions
 //===----------------------------------------------------------------------===//
 
 def MOPPredicate : ScalableVectorOfLengthAndType<[16, 8, 4, 2], [I1]>;
 def MOPVector : ScalableVectorOfLengthAndType<[16, 8, 4, 2],
                                               [I8, I16, BF16, F16, F32, F64]>;
 def LDSTPredicate : ScalableVectorOfLengthAndType<[16, 8, 4, 2, 1], [I1]>;

mlir/lib/Conversion/VectorToArmSME/VectorToArmSME.cpp

 };
 
 /// Conversion pattern for dense arith.constant.
 struct ConstantOpToArmSMELowering : public OpRewritePattern<arith::ConstantOp> {
   using OpRewritePattern<arith::ConstantOp>::OpRewritePattern;
 
   LogicalResult matchAndRewrite(arith::ConstantOp constantOp,
                                 PatternRewriter &rewriter) const final {
-    auto vType = dyn_cast<VectorType>(constantOp.getType());
-    if (!vType || !arm_sme::isValidSMETileVectorType(vType))
+    auto tileType = dyn_cast<VectorType>(constantOp.getType());
+    if (!tileType || !arm_sme::isValidSMETileVectorType(tileType))
       return failure();
 
     auto denseAttr = dyn_cast<DenseElementsAttr>(constantOp.getValueAttr());
-    if (!denseAttr || !isSplatZero(vType.getElementType(), denseAttr))
+    if (!denseAttr || !denseAttr.isSplat())
       return failure();
 
-    rewriter.replaceOpWithNewOp<arm_sme::ZeroOp>(constantOp, vType);
+    auto tileElementType = tileType.getElementType();
+
+    // Lower 'arith.constant dense<0>' to 'arm_sme.zero' op.
+    if (isSplatZero(tileElementType, denseAttr)) {
+      rewriter.replaceOpWithNewOp<arm_sme::ZeroOp>(constantOp, tileType);
+      return success();
+    }
+
+    // Lower non-zero constants to a loop of 'arm_sme.move_vector_to_tile_slice'
+    // ops that broadcast the constant to each tile slice.
+    OpBuilder::InsertionGuard g(rewriter);
+    auto loc = constantOp.getLoc();
+
+    // Unpack 1-d vector type from 2-d vector type.
+    auto tileSliceType =
+        VectorType::get(tileType.getShape().drop_front(), tileElementType,
+                        /*scalableDims=*/{true});
+    auto denseAttr1D = DenseElementsAttr::get(
+        tileSliceType, denseAttr.getSplatValue<Attribute>());
+    auto constantOp1D = rewriter.create<arith::ConstantOp>(loc, denseAttr1D);
+
+    unsigned tileElementWidth = tileElementType.getIntOrFloatBitWidth();
+
+    // Create 'arm_sme.get_tile' op.
+    auto tileId = rewriter.create<arm_sme::GetTileID>(
+        loc, rewriter.getIntegerType(tileElementWidth));
+
+    // Create `arm_sme.cast_tile_to_vector` to cast tile ID to a vector type to
+    // use as input tile to 'arm_sme.move_vector_to_tile_slice' ops.
+    auto tile =
+        rewriter.create<arm_sme::CastTileToVector>(loc, tileType, tileId);
+
+    auto step = rewriter.create<arith::ConstantIndexOp>(loc, 1);
+    auto minTileSlices = rewriter.create<arith::ConstantIndexOp>(
+        loc, arm_sme::getSMETileSliceMinNumElts(tileElementType));
+    auto vscale =
+        rewriter.create<vector::VectorScaleOp>(loc, rewriter.getIndexType());
+    auto lowerBound = rewriter.create<arith::ConstantIndexOp>(loc, 0);
+    auto numTileSlices =
+        rewriter.create<arith::MulIOp>(loc, minTileSlices, vscale);
+    // Create a loop that broadcasts the constant to each ZA tile slice.
+    auto forOp =
+        rewriter.create<scf::ForOp>(loc, lowerBound, numTileSlices, step);
+    rewriter.setInsertionPointToStart(forOp.getBody());
+    auto tileSliceIndex = forOp.getInductionVar();
+
+    // Create 'arm_sme.move_vector_to_tile_slice' to write vector to tile slice.
+    rewriter.create<arm_sme::MoveVectorToTileSliceOp>(
+        loc, tileType, constantOp1D, tile, tileSliceIndex);
+
+    rewriter.setInsertionPointAfter(forOp);
+
+    rewriter.replaceOp(constantOp, tile);
 
     return success();
   }
 };
 
 } // namespace
 
 void mlir::populateVectorToArmSMEPatterns(RewritePatternSet &patterns,
                                           MLIRContext &ctx) {
   patterns.add<TransferWriteToArmSMELowering, VectorLoadToArmSMELowering,
                VectorStoreToArmSMELowering, ConstantOpToArmSMELowering>(&ctx);
 }

mlir/lib/Dialect/ArmSME/Transforms/LegalizeForLLVMExport.cpp

case 128:
           storeTileSliceOp, allActiveMask, ptr, tileI32, tileSliceI32);
       break;
     }
 
     return success();
   }
 };
 
+/// Lower `arm_sme.move_vector_to_tile_slice` to SME intrinsics. Only horizontal
+/// tile slices are currently supported.
+struct MoveVectorToTileSliceToArmSMELowering
+    : public ConvertOpToLLVMPattern<arm_sme::MoveVectorToTileSliceOp> {
+  using ConvertOpToLLVMPattern<
+      arm_sme::MoveVectorToTileSliceOp>::ConvertOpToLLVMPattern;
+
+  LogicalResult
+  matchAndRewrite(arm_sme::MoveVectorToTileSliceOp moveVectorToTileSliceOp,
+                  arm_sme::MoveVectorToTileSliceOp::Adaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    auto loc = moveVectorToTileSliceOp.getLoc();
+    auto tileType = moveVectorToTileSliceOp.getTileType();
+    auto tileElementType = tileType.getElementType();
+    unsigned tileElementWidth = tileElementType.getIntOrFloatBitWidth();
+
+    // Create 'arm_sme.cast_vector_to_tile' to get a tile ID for the tile being
+    // loaded to.
+    auto tile = rewriter.create<arm_sme::CastVectorToTile>(
+        loc, rewriter.getIntegerType(tileElementWidth),
+        moveVectorToTileSliceOp.getTile());
+
+    auto tileSlice = moveVectorToTileSliceOp.getTileSliceIndex();
+
+    // Cast tile slice from index to i32 for intrinsic.
+    auto tileSliceI32 = rewriter.create<arith::IndexCastUIOp>(
+        loc, rewriter.getI32Type(), tileSlice);
+
+    // Create all active predicate mask.
+    auto one = rewriter.create<arith::ConstantOp>(
+        loc, rewriter.getI1Type(),
+        rewriter.getIntegerAttr(rewriter.getI1Type(), 1));
+    auto predTy = VectorType::get(tileType.getShape()[0], rewriter.getI1Type(),
+                                  /*scalableDims=*/{true});
+    auto allActiveMask = rewriter.create<vector::SplatOp>(loc, predTy, one);
+
+    auto tileI32 = castTileIDToI32(tile, loc, rewriter);
+
+    // Create 'arm_sme.intr.write.horiz' to write vector to tile slice.
+    rewriter.create<arm_sme::aarch64_sme_write_horiz>(
+        loc, tileI32, tileSliceI32, allActiveMask,
+        moveVectorToTileSliceOp.getVector());
+
+    // Intrinsic has no result, replace 'arm_sme.move_vector_to_tile_slice' with
+    // 'arm_sme.cast_tile_to_vector' to preserve dataflow.
+    rewriter.replaceOpWithNewOp<arm_sme::CastTileToVector>(
+        moveVectorToTileSliceOp, tileType, tile);
+
+    return success();
+  }
+};
+
 } // namespace
 
 void mlir::configureArmSMELegalizeForExportTarget(
     LLVMConversionTarget &target) {
   target.addLegalOp<
       scf::ForOp, scf::YieldOp, arm_sme::CastTileToVector,
       arm_sme::CastVectorToTile, arm_sme::aarch64_sme_zero,
       arm_sme::aarch64_sme_str, arm_sme::aarch64_sme_ld1b_horiz,
       arm_sme::aarch64_sme_ld1h_horiz, arm_sme::aarch64_sme_ld1w_horiz,
       arm_sme::aarch64_sme_ld1d_horiz, arm_sme::aarch64_sme_ld1q_horiz,
       arm_sme::aarch64_sme_st1b_horiz, arm_sme::aarch64_sme_st1h_horiz,
       arm_sme::aarch64_sme_st1w_horiz, arm_sme::aarch64_sme_st1d_horiz,
-      arm_sme::aarch64_sme_st1q_horiz, arm_sme::aarch64_sme_za_enable,
-      arm_sme::aarch64_sme_za_disable>();
+      arm_sme::aarch64_sme_st1q_horiz, arm_sme::aarch64_sme_write_horiz,
+      arm_sme::aarch64_sme_za_enable, arm_sme::aarch64_sme_za_disable>();
   target.addLegalOp<GetTileID>();
 
   // Mark 'func.func' ops as legal if either:
   //   1. no 'arm_za' function attribute is present.
   //   2. the 'arm_za' function attribute is present and the first op in the
   //      function is an 'arm_sme::aarch64_sme_za_enable' intrinsic.
   target.addDynamicallyLegalOp<func::FuncOp>([&](func::FuncOp funcOp) {
     if (funcOp.isDeclaration())
target.addDynamicallyLegalOp<func::ReturnOp>([&](func::ReturnOp returnOp) {
     return !funcOp->hasAttr("arm_za") || hasDisableZA;
   });
 }
 
 void mlir::populateArmSMELegalizeForLLVMExportPatterns(
     LLVMTypeConverter &converter, RewritePatternSet &patterns) {
   patterns.add<EnableZAPattern, DisableZAPattern>(patterns.getContext());
   patterns.add<ZeroOpConversion, StoreTileSliceToArmSMELowering,
-               LoadTileSliceToArmSMELowering>(converter);
+               LoadTileSliceToArmSMELowering,
+               MoveVectorToTileSliceToArmSMELowering>(converter);
 }

mlir/test/Dialect/ArmSME/arith-ops-to-sme.mlir

 
 // CHECK-LABEL: @arith_constant_dense_2d_zero_f64
 // CHECK: %[[ZERO:.*]] = arm_sme.zero : vector<[2]x[2]xf64>
 func.func @arith_constant_dense_2d_zero_f64() {
   %zero = arith.constant dense<0.0> : vector<[2]x[2]xf64>
   "prevent.dce"(%zero) : (vector<[2]x[2]xf64>) -> ()
   return
 }
+
+// =============================================================================
+// Non-zero arith.constant dense to SME
+// =============================================================================
+
+// -----
+
+// CHECK-LABEL: func.func @arith_constant_dense_2d_nonzero_i8() {
+// CHECK: %[[C2_SPLAT:.*]] = arith.constant dense<2> : vector<[16]xi8>
+// CHECK: %[[C1:.*]] = arith.constant 1 : index
+// CHECK: %[[C16:.*]] = arith.constant 16 : index
+// CHECK: %[[C0:.*]] = arith.constant 0 : index
+// CHECK: %[[GET_TILE_ID:.*]] = arm_sme.get_tile_id : i8
+// CHECK: %[[TILE:.*]] = arm_sme.cast_tile_to_vector %[[GET_TILE_ID]] : i8 to vector<[16]x[16]xi8>
+// CHECK: %[[VSCALE:.*]] = vector.vscale
+// CHECK: %[[NUM_TILE_SLICES:.*]] = arith.muli %[[VSCALE]], %[[C16]] : index
+// CHECK: scf.for %[[TILE_SLICE_INDEX:.*]] = %[[C0]] to %[[NUM_TILE_SLICES]] step %[[C1]] {
+// CHECK:   arm_sme.move_vector_to_tile_slice %[[C2_SPLAT]], %[[TILE]], %[[TILE_SLICE_INDEX]] : vector<[16]xi8> into vector<[16]x[16]xi8>
+// CHECK: "prevent.dce"(%[[TILE]]) : (vector<[16]x[16]xi8>) -> ()
+func.func @arith_constant_dense_2d_nonzero_i8() {
+  %two = arith.constant dense<2> : vector<[16]x[16]xi8>
+  "prevent.dce"(%two) : (vector<[16]x[16]xi8>) -> ()
+  return
+}
+
+// -----
+
+// CHECK-LABEL: func.func @arith_constant_dense_2d_nonzero_f64() {
+// CHECK: %[[C2_SPLAT:.*]] = arith.constant dense<2.000000e+00> : vector<[2]xf64>
+// CHECK: %[[C1:.*]] = arith.constant 1 : index
+// CHECK: %[[C2:.*]] = arith.constant 2 : index
+// CHECK: %[[C0:.*]] = arith.constant 0 : index
+// CHECK: %[[GET_TILE_ID:.*]] = arm_sme.get_tile_id : i64
+// CHECK: %[[TILE:.*]] = arm_sme.cast_tile_to_vector %[[GET_TILE_ID]] : i64 to vector<[2]x[2]xf64>
+// CHECK: %[[VSCALE:.*]] = vector.vscale
+// CHECK: %[[NUM_TILE_SLICES:.*]] = arith.muli %[[VSCALE]], %[[C2]] : index
+// CHECK: scf.for %[[TILE_SLICE_INDEX:.*]] = %[[C0]] to %[[NUM_TILE_SLICES]] step %[[C1]] {
+// CHECK:   arm_sme.move_vector_to_tile_slice %[[C2_SPLAT]], %[[TILE]], %[[TILE_SLICE_INDEX]] : vector<[2]xf64> into vector<[2]x[2]xf64>
+// CHECK: "prevent.dce"(%[[TILE]]) : (vector<[2]x[2]xf64>) -> ()
+func.func @arith_constant_dense_2d_nonzero_f64() {
+  %two = arith.constant dense<2.0> : vector<[2]x[2]xf64>
+  "prevent.dce"(%two) : (vector<[2]x[2]xf64>) -> ()
+  return
+}

mlir/test/Dialect/ArmSME/invalid.mlir

 
 // -----
 
 func.func @arm_sme_get_tile_id__bad_type() -> i1 {
   // expected-error@+1 {{op result #0 must be 8-bit signless integer or 16-bit signless integer or 32-bit signless integer or 64-bit signless integer or 128-bit signless integer}}
   %0 = arm_sme.get_tile_id : i1
   return %0 : i1
 }
+
+// -----
+
+func.func @arm_sme_move_vector_to_tile_slice_i8__bad_vector_type(%vector : vector<[8]xi8>, %tile : vector<[16]x[16]xi8>, %tile_slice_index : index) -> vector<[16]x[16]xi8> {
+  %c0 = arith.constant 0 : index
+  // expected-error@+1 {{op failed to verify that type of 'vector' matches type of 'tile' slice}}
+  %0 = arm_sme.move_vector_to_tile_slice %vector, %tile, %tile_slice_index : vector<[8]xi8> into vector<[16]x[16]xi8>
+  return %0 : vector<[16]x[16]xi8>
+}
+
+// -----
+
+func.func @arm_sme_move_vector_to_tile_slice_f32__bad_vector_type(%vector : vector<[8]xf32>, %tile : vector<[4]x[4]xf32>, %tile_slice_index : index) -> vector<[4]x[4]xf32> {
+  %c0 = arith.constant 0 : index
+  // expected-error@+1 {{op failed to verify that type of 'vector' matches type of 'tile' slice}}
+  %0 = arm_sme.move_vector_to_tile_slice %vector, %tile, %tile_slice_index : vector<[8]xf32> into vector<[4]x[4]xf32>
+  return %0 : vector<[4]x[4]xf32>
+}

mlir/test/Dialect/ArmSME/roundtrip.mlir

 // -----
 
 func.func @arm_sme_store_tile_slice_f64(%tile : vector<[2]x[2]xf64>, %tile_slice_index : index, %dest : memref<?x?xf64>) -> () {
   // CHECK: arm_sme.store_tile_slice {{.*}} : memref<?x?xf64>, vector<[2]x[2]xf64>
   %c0 = arith.constant 0 : index
   arm_sme.store_tile_slice %tile, %tile_slice_index, %dest[%c0] : memref<?x?xf64>, vector<[2]x[2]xf64>
   return
 }
+
+// -----
+
+func.func @arm_sme_move_vector_to_tile_slice_i8(%vector : vector<[16]xi8>, %tile : vector<[16]x[16]xi8>, %tile_slice_index : index) -> () {
+  // CHECK: arm_sme.move_vector_to_tile_slice {{.*}} : vector<[16]xi8> into vector<[16]x[16]xi8>
+  %c0 = arith.constant 0 : index
+  arm_sme.move_vector_to_tile_slice %vector, %tile, %tile_slice_index : vector<[16]xi8> into vector<[16]x[16]xi8>
+  return
+}
+
+// -----
+
+func.func @arm_sme_move_vector_to_tile_slice_i16(%vector : vector<[8]xi16>, %tile : vector<[8]x[8]xi16>, %tile_slice_index : index) -> () {
+  // CHECK: arm_sme.move_vector_to_tile_slice {{.*}} : vector<[8]xi16> into vector<[8]x[8]xi16>
+  %c0 = arith.constant 0 : index
+  arm_sme.move_vector_to_tile_slice %vector, %tile, %tile_slice_index : vector<[8]xi16> into vector<[8]x[8]xi16>
+  return
+}
+
+// -----
+
+func.func @arm_sme_move_vector_to_tile_slice_i32(%vector : vector<[4]xi32>, %tile : vector<[4]x[4]xi32>, %tile_slice_index : index) -> () {
+  // CHECK: arm_sme.move_vector_to_tile_slice {{.*}} : vector<[4]xi32> into vector<[4]x[4]xi32>
+  %c0 = arith.constant 0 : index
+  arm_sme.move_vector_to_tile_slice %vector, %tile, %tile_slice_index : vector<[4]xi32> into vector<[4]x[4]xi32>
+  return
+}
+
+// -----
+
+func.func @arm_sme_move_vector_to_tile_slice_i64(%vector : vector<[2]xi64>, %tile : vector<[2]x[2]xi64>, %tile_slice_index : index) -> () {
+  // CHECK: arm_sme.move_vector_to_tile_slice {{.*}} : vector<[2]xi64> into vector<[2]x[2]xi64>
+  %c0 = arith.constant 0 : index
+  arm_sme.move_vector_to_tile_slice %vector, %tile, %tile_slice_index : vector<[2]xi64> into vector<[2]x[2]xi64>
+  return
+}
+
+// -----
+
+func.func @arm_sme_move_vector_to_tile_slice_i128(%vector : vector<[1]xi128>, %tile : vector<[1]x[1]xi128>, %tile_slice_index : index) -> () {
+  // CHECK: arm_sme.move_vector_to_tile_slice {{.*}} : vector<[1]xi128> into vector<[1]x[1]xi128>
+  %c0 = arith.constant 0 : index
+  arm_sme.move_vector_to_tile_slice %vector, %tile, %tile_slice_index : vector<[1]xi128> into vector<[1]x[1]xi128>
+  return
+}
+
+// -----
+
+func.func @arm_sme_move_vector_to_tile_slice_f16(%vector : vector<[8]xf16>, %tile : vector<[8]x[8]xf16>, %tile_slice_index : index) -> () {
+  // CHECK: arm_sme.move_vector_to_tile_slice {{.*}} : vector<[8]xf16> into vector<[8]x[8]xf16>
+  %c0 = arith.constant 0 : index
+  arm_sme.move_vector_to_tile_slice %vector, %tile, %tile_slice_index : vector<[8]xf16> into vector<[8]x[8]xf16>
+  return
+}
+
+// -----
+
+func.func @arm_sme_move_vector_to_tile_slice_bf16(%vector : vector<[8]xbf16>, %tile : vector<[8]x[8]xbf16>, %tile_slice_index : index) -> () {
+  // CHECK: arm_sme.move_vector_to_tile_slice {{.*}} : vector<[8]xbf16> into vector<[8]x[8]xbf16>
+  %c0 = arith.constant 0 : index
+  arm_sme.move_vector_to_tile_slice %vector, %tile, %tile_slice_index : vector<[8]xbf16> into vector<[8]x[8]xbf16>
+  return
+}
+
+// -----
+
+func.func @arm_sme_move_vector_to_tile_slice_f32(%vector : vector<[4]xf32>, %tile : vector<[4]x[4]xf32>, %tile_slice_index : index) -> () {
+  // CHECK: arm_sme.move_vector_to_tile_slice {{.*}} : vector<[4]xf32> into vector<[4]x[4]xf32>
+  %c0 = arith.constant 0 : index
+  arm_sme.move_vector_to_tile_slice %vector, %tile, %tile_slice_index : vector<[4]xf32> into vector<[4]x[4]xf32>
+  return
+}
+
+// -----
+
+func.func @arm_sme_move_vector_to_tile_slice_f64(%vector : vector<[2]xf64>, %tile : vector<[2]x[2]xf64>, %tile_slice_index : index) -> () {
+  // CHECK: arm_sme.move_vector_to_tile_slice {{.*}} : vector<[2]xf64> into vector<[2]x[2]xf64>
+  %c0 = arith.constant 0 : index
+  arm_sme.move_vector_to_tile_slice %vector, %tile, %tile_slice_index : vector<[2]xf64> into vector<[2]x[2]xf64>
+  return
+}

mlir/test/Integration/Dialect/Vector/CPU/ArmSME/tile_fill.mlir

+// RUN: mlir-opt %s -enable-arm-streaming="mode=locally enable-za" \
+// RUN:   -convert-vector-to-arm-sme -convert-arm-sme-to-scf \
+// RUN:   -convert-vector-to-llvm="enable-arm-sme" -cse -canonicalize \
+// RUN:   -allocate-arm-sme-tiles -test-lower-to-llvm | \
+// RUN: %mcr_aarch64_cmd \
+// RUN:  -march=aarch64 -mattr=+sve,+sme \
+// RUN:  -e entry -entry-point-result=i32 \
+// RUN:  -shared-libs=%mlir_runner_utils,%mlir_c_runner_utils | \
+// RUN: FileCheck %s
+
+// Integration test demonstrating filling a 32-bit element ZA tile with a
+// non-zero constant via vector to tile (MOVA) ops.
+
+llvm.func @printCString(!llvm.ptr<i8>)
+
+func.func @printTileBegin() {
+  %0 = llvm.mlir.addressof @str_tile_begin : !llvm.ptr<array<11 x i8>>
+  %1 = llvm.mlir.constant(0 : index) : i64
+  %2 = llvm.getelementptr %0[%1, %1]
+    : (!llvm.ptr<array<11 x i8>>, i64, i64) -> !llvm.ptr<i8>
+  llvm.call @printCString(%2) : (!llvm.ptr<i8>) -> ()
+  return
+}
+
+func.func @printTileEnd() {
+  %0 = llvm.mlir.addressof @str_tile_end : !llvm.ptr<array<9 x i8>>
+  %1 = llvm.mlir.constant(0 : index) : i64
+  %2 = llvm.getelementptr %0[%1, %1]
+    : (!llvm.ptr<array<9 x i8>>, i64, i64) -> !llvm.ptr<i8>
+  llvm.call @printCString(%2) : (!llvm.ptr<i8>) -> ()
+  return
+}
+
+func.func @entry() -> i32 {
+  %c0 = arith.constant 0 : index
+  %c1_index = arith.constant 1 : index
+
+  %min_elts_s = arith.constant 4 : index
+  %vscale = vector.vscale
+
+  // "svl" refers to the Streaming Vector Length and "svl_s" the number of
+  // 32-bit elements in a vector of SVL bits.
+  %svl_s = arith.muli %min_elts_s, %vscale : index
+
+  // Allocate memory.
+  %tilesize = arith.muli %svl_s, %svl_s : index
+  %mem = memref.alloca(%tilesize) : memref<?xi32>
+
+  // Fill a tile with '123'. This will get lowered to a 1-d vector splat of
+  // '123' and a loop that writes this vector to each tile slice in the ZA
+  // tile.
+  %tile = arith.constant dense<123> : vector<[4]x[4]xi32>
+
+  // Store tile to memory so it can be dumped.
+  vector.store %tile, %mem[%c0] : memref<?xi32>, vector<[4]x[4]xi32>
+
+  // Dump "mem". The smallest SVL is 128-bits so the tile will be at least
+  // 4x4xi32.
+  //
+  // CHECK:      TILE BEGIN
+  // CHECK-NEXT: ( 123, 123, 123, 123
+  // CHECK-NEXT: ( 123, 123, 123, 123
+  // CHECK-NEXT: ( 123, 123, 123, 123
+  // CHECK-NEXT: ( 123, 123, 123, 123
+  // CHECK:      TILE END
+  func.call @printTileBegin() : () -> ()
+  scf.for %i = %c0 to %tilesize step %svl_s {
+    %tileslice = vector.load %mem[%i] : memref<?xi32>, vector<[4]xi32>
+    vector.print %tileslice : vector<[4]xi32>
+  }
+  func.call @printTileEnd() : () -> ()
+
+  %c0_i32 = arith.constant 0 : i32
+  return %c0_i32 : i32
+}
+
+llvm.mlir.global internal constant @str_tile_begin("TILE BEGIN\0A")
+llvm.mlir.global internal constant @str_tile_end("TILE END\0A")