[mlir][ArmSVE] Add convert_to/from_svbool ops

mlir/include/mlir/Dialect/ArmSVE/IR/ArmSVE.td

@@ -28,6 +28,8 @@ def ArmSVE_Dialect : Dialect {
     This dialect contains the definitions necessary to target specific Arm SVE
     scalable vector operations.
   }];
+
+  let dependentDialects = ["vector::VectorDialect"];
 }
 
 //===----------------------------------------------------------------------===//
@@ -40,6 +42,11 @@ def SVBool : ScalableVectorOfRankAndLengthAndType<
 def SVEPredicate : ScalableVectorOfRankAndLengthAndType<
   [1], [16, 8, 4, 2, 1], [I1]>;
 
+// Generalizations of SVBool and SVEPredicate to ranks >= 1.
+// These are masks with a single trailing scalable dimension.
+def SVBoolMask : TrailingScalableVectorOfSizeAndType<[16], [I1]>;
+def SVEPredicateMask : TrailingScalableVectorOfSizeAndType<[16, 8, 4, 2, 1], [I1]>;
+
 //===----------------------------------------------------------------------===//
 // ArmSVE op definitions
 //===----------------------------------------------------------------------===//
@@ -236,6 +243,81 @@ def UmmlaOp : ArmSVE_Op<"ummla",
     "$acc `,` $src1 `,` $src2 attr-dict `:` type($src1) `to` type($dst)";
 }
 
+class SvboolTypeConstraint<string lhsArg, string rhsArg> : TypesMatchWith<
+      "expected corresponding svbool type widened to [16]xi1",
+      lhsArg, rhsArg,
+      "VectorType(VectorType::Builder(::llvm::cast<VectorType>($_self)).setDim(::llvm::cast<VectorType>($_self).getRank() - 1, 16))">;
+
+def ConvertFromSvboolOp : ArmSVE_Op<"convert_from_svbool",
+                            [Pure, SvboolTypeConstraint<"result", "source">]>
+{
+  let summary = "Convert a svbool type to a SVE predicate type";
+  let description = [{
+    Converts svbool types (`vector<[16]xi1>` or vectors of that type, e.g.
+    `vector<2x3x[16]xi1>`) to SVE predicate types. Note: Only the trailing
+    dimension can be scalable.
+
+    Example 1: Convert a 1-D svbool mask to a SVE predicate.
+    ```mlir
+    %source = vector.load %memref[%c0] : memref<?xi1>, vector<[16]xi1>
+    %result = arm_sve.convert_from_svbool %source : vector<[4]xi1>
+    ```
+
+    Example 2: Convert a 2-D svbool mask to a mask of SVE predicates.
+    ```mlir
+    %source = vector.load %memref[%c0, %c0] : memref<2x?xi1>, vector<2x[16]xi1>
+    %result = arm_sve.convert_from_svbool %source : vector<2x[8]xi1>
+    ```
+
+    ---
+
+    A `svbool` is the smallest SVE predicate type that has a in-memory
+    representation (and maps to a full predicate register). In MLIR `svbool` is
+    represented as `vector<[16]xi1>`. Smaller SVE predicate types
+    (`vector<[1|2|4|8]xi1>`) must be stored as `svbool` then converted back to
+    a predicate after loading.
+  }];
+  let arguments = (ins SVBoolMask:$source);
+  let results = (outs SVEPredicateMask:$result);
+  let assemblyFormat = "$source attr-dict `:` type($result)";
+}
+
+def ConvertToSvboolOp : ArmSVE_Op<"convert_to_svbool",
+                            [Pure, SvboolTypeConstraint<"source", "result">]>
+{
+  let summary = "Convert a SVE predicate type to a svbool type";
+  let description = [{
+    Converts SVE predicate types (or vectors of predicate types, e.g.
+    `vector<4x[4]xi1>`) to svbool types. Note: Only the trailing dimension can
+    be scalable.
+
+    Example 1: Convert a 1-D SVE predicate to a svbool mask.
+    ```mlir
+    %source = vector.create_mask %dim_size : vector<[4]xi1>
+    %result = arm_sve.convert_to_svbool %source : vector<[4]xi1>
+    // => Results in vector<[16]xi1>
+    ```
+
+    Example 2: Convert a 2-D mask of SVE predicates to a svbool mask.
+    ```mlir
+    %source = vector.create_mask %c2, %dim_size : vector<2x[2]xi1>
+    %result = arm_sve.convert_to_svbool %source : vector<2x[2]xi1>
+    // => Results in vector<2x[16]xi1>
+    ```
+
+    ---
+
+    A `svbool` is the smallest SVE predicate type that has a in-memory
+    representation (and maps to a full predicate register). In MLIR `svbool` is
+    represented as `vector<[16]xi1>`. Smaller SVE predicate types
+    (`vector<[1|2|4|8]xi1>`) must be converted to a `svbool` before they can be
+    stored.
+  }];
+  let arguments = (ins SVEPredicateMask:$source);
+  let results = (outs SVBoolMask:$result);
+  let assemblyFormat = "$source attr-dict `:` type($source)";
+}
+
 def ScalableMaskedAddIOp : ScalableMaskedIOp<"masked.addi", "addition",
                                              [Commutative]>;
 

mlir/include/mlir/IR/CommonTypeConstraints.td

@@ -37,6 +37,19 @@ def IsFixedVectorTypePred : CPred<[{::llvm::isa<::mlir::VectorType>($_self) &&
 def IsScalableVectorTypePred : CPred<[{::llvm::isa<::mlir::VectorType>($_self) &&
                                    ::llvm::cast<VectorType>($_self).isScalable()}]>;
 
+// Whether a type is a scalable VectorType, with a single trailing scalable dimension.
+// Examples:
+// Valid:
+//   - vector<[4]xf32>, vector<2x3x[2]xi64>, vector<32x[8]xi32>
+// Invalid
+//   - vector<[4]x8xi32>, vector<[2]x[2]xf64>, vector<2x[8]x4xi32>
+def IsOnlyTrailingDimScalablePred : And<[
+  CPred<"::llvm::isa<::mlir::VectorType>($_self)">,
+  CPred<"::llvm::cast<::mlir::VectorType>($_self).getRank() > 0">,
+  CPred<"::llvm::cast<::mlir::VectorType>($_self).getScalableDims().back()">,
+  CPred<"!llvm::is_contained(::llvm::cast<::mlir::VectorType>($_self).getScalableDims().drop_back(), true)">
+]>;
+
 // Whether a type is a VectorType and all dimensions are scalable.
 def allDimsScalableVectorTypePred : And<[
   IsVectorTypePred,
@@ -404,6 +417,12 @@ class ScalableVectorOf<list<Type> allowedTypes> :
   ShapedContainerType<allowedTypes, IsScalableVectorTypePred,
           "scalable vector", "::mlir::VectorType">;
 
+// Any vector with a single trailing scalable dimension, with an element type in
+// the `allowedTypes` list.
+class TrailingScalableVectorOf<list<Type> allowedTypes> :
+  ShapedContainerType<allowedTypes, IsOnlyTrailingDimScalablePred,
+          "trailing scalable vector", "::mlir::VectorType">;
+
 // Whether the number of elements of a vector is from the given
 // `allowedRanks` list
 class IsVectorOfRankPred<list<int> allowedRanks> :
@@ -481,10 +500,32 @@ class IsScalableVectorOfLengthPred<list<int> allowedLengths> :
                            == }]
                          # allowedlength>)>]>;
 
+class abs<int value> {
+  int ret = !if(!lt(value, 0), !sub(0, value), value);
+}
+
+// Whether the n-th (starting from 1) dim of the shape matches the given `size`.
+// Negative values index in reverse.
+class IsNthDimSizeIsOneOfPred<int n, list<int> allowedSizes>
+  : And<[CPred<"::llvm::cast<::mlir::ShapedType>($_self).getRank() >= " # abs<n>.ret>,
+        CPred<"::llvm::is_contained(ArrayRef<int64_t>({" # !interleave(allowedSizes, ", ") # "}), "
+          # "::llvm::cast<::mlir::ShapedType>($_self).getDimSize("
+          #   !if(!lt(n, 0),
+                "::llvm::cast<::mlir::ShapedType>($_self).getRank() + " # n,
+                "" # !sub(n, 1))
+          # "))">]>;
+
 // Whether the shape of a vector matches the given `shape` list.
 class IsVectorOfShape<list<int> shape>
   : CPred<"::llvm::cast<::mlir::VectorType>($_self).getShape() == ArrayRef<int64_t>({" # !interleave(shape, ", ") # "})">;
 
+// Any ShapedType where the size of the n-th dim is contained in `sizes`.
+// Negative values index in reverse.
+class ShapedTypeWithNthDimOfSize<int n, list<int> allowedSizes> : Type<
+  IsNthDimSizeIsOneOfPred<n, allowedSizes>,
+  " with dim " # n # " having a size of {" # !interleave(allowedSizes, ", ") # "}",
+  "::mlir::ShapedType">;
+
 // Any vector where the number of elements is from the given
 // `allowedLengths` list
 class VectorOfLength<list<int> allowedLengths> : Type<
@@ -546,6 +587,17 @@ class ScalableVectorOfRankAndLengthAndType<list<int> allowedRanks,
   ScalableVectorOfLength<allowedLengths>.summary,
   "::mlir::VectorType">;
 
+// Any scalable vector with a single trailing scalable dimensions, where the
+// size of the trailing dimension is in `allowedTrailingSizes` list, and the
+// type is in the `allowedTypes` list.
+class TrailingScalableVectorOfSizeAndType<list<int> allowedTrailingSizes,
+                                           list<Type> allowedTypes> : AllOfType<
+  [TrailingScalableVectorOf<allowedTypes>,
+   ShapedTypeWithNthDimOfSize<-1, allowedTrailingSizes>],
+   TrailingScalableVectorOf<allowedTypes>.summary #
+   ShapedTypeWithNthDimOfSize<-1, allowedTrailingSizes>.summary,
+  "::mlir::VectorType">;
+
 def AnyVector : VectorOf<[AnyType]>;
 // Temporary vector type clone that allows gradual transition to 0-D vectors.
 def AnyVectorOfAnyRank : VectorOfAnyRankOf<[AnyType]>;

mlir/lib/Dialect/ArmSVE/IR/ArmSVEDialect.cpp

@@ -12,6 +12,7 @@
 
 #include "mlir/Dialect/ArmSVE/IR/ArmSVEDialect.h"
 #include "mlir/Dialect/LLVMIR/LLVMTypes.h"
+#include "mlir/Dialect/Vector/IR/VectorOps.h"
 #include "mlir/IR/Builders.h"
 #include "mlir/IR/DialectImplementation.h"
 #include "mlir/IR/OpImplementation.h"

mlir/lib/Dialect/ArmSVE/IR/CMakeLists.txt

@@ -10,5 +10,6 @@ add_mlir_dialect_library(MLIRArmSVEDialect
   LINK_LIBS PUBLIC
   MLIRIR
   MLIRLLVMDialect
+  MLIRVectorDialect
   MLIRSideEffectInterfaces
   )

mlir/lib/Dialect/ArmSVE/Transforms/CMakeLists.txt

@@ -7,6 +7,7 @@ add_mlir_dialect_library(MLIRArmSVETransforms
   LINK_LIBS PUBLIC
   MLIRArmSVEDialect
   MLIRFuncDialect
+  MLIRVectorDialect
   MLIRIR
   MLIRLLVMCommonConversion
   MLIRLLVMDialect

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

@@ -12,6 +12,8 @@
 #include "mlir/Dialect/ArmSVE/Transforms/Transforms.h"
 #include "mlir/Dialect/Func/IR/FuncOps.h"
 #include "mlir/Dialect/LLVMIR/LLVMDialect.h"
+#include "mlir/Dialect/Utils/IndexingUtils.h"
+#include "mlir/Dialect/Vector/IR/VectorOps.h"
 #include "mlir/IR/BuiltinOps.h"
 #include "mlir/IR/PatternMatch.h"
 
@@ -66,6 +68,77 @@ using ScalableMaskedDivFOpLowering =
     OneToOneConvertToLLVMPattern<ScalableMaskedDivFOp,
                                  ScalableMaskedDivFIntrOp>;
 
+namespace {
+
+/// Unrolls a conversion to/from equivalent vector types, to allow using a
+/// conversion intrinsic that only supports 1-D vector types.
+///
+/// Example:
+/// ```
+/// %result = arm_sve.convert_to_svbool %source : vector<2x[4]xi1>
+/// ```
+/// is rewritten into:
+/// ```
+/// %cst = arith.constant dense<false> : vector<2x[16]xi1>
+/// %1 = vector.extract %source[0] : vector<[4]xi1> from vector<2x[4]xi1>
+/// %2 = "arm_sve.intr.convert.to.svbool"(%1)
+///                : (vector<[4]xi1>) -> vector<[16]xi1>
+/// %3 = vector.insert %2, %cst [0] : vector<[16]xi1> into vector<2x[16]xi1>
+/// %4 = vector.extract %source[1] : vector<[4]xi1> from vector<2x[4]xi1>
+/// %5 = "arm_sve.intr.convert.to.svbool"(%4)
+///                : (vector<[4]xi1>) -> vector<[16]xi1>
+/// %result = vector.insert %5, %3 [1] : vector<[16]xi1> into vector<2x[16]xi1>
+/// ```
+template <typename Op, typename IntrOp>
+struct SvboolConversionOpLowering : public ConvertOpToLLVMPattern<Op> {
+  using ConvertOpToLLVMPattern<Op>::ConvertOpToLLVMPattern;
+
+  LogicalResult
+  matchAndRewrite(Op convertOp, typename Op::Adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    auto loc = convertOp.getLoc();
+
+    auto source = convertOp.getSource();
+    VectorType sourceType = source.getType();
+    VectorType resultType = convertOp.getResult().getType();
+
+    Value result = rewriter.create<arith::ConstantOp>(
+        loc, resultType, rewriter.getZeroAttr(resultType));
+
+    // We want to iterate over the input vector in steps of the trailing
+    // dimension. So this creates tile shape where all leading dimensions are 1,
+    // and the trailing dimension step is the size of the dimension.
+    SmallVector<int64_t> tileShape(sourceType.getRank(), 1);
+    tileShape.back() = sourceType.getShape().back();
+
+    // Iterate over all scalable mask/predicate slices of the source vector.
+    for (SmallVector<int64_t> index :
+         StaticTileOffsetRange(sourceType.getShape(), tileShape)) {
+      auto extractOrInsertPosition = ArrayRef(index).drop_back();
+      auto sourceVector = rewriter.create<vector::ExtractOp>(
+          loc, source, extractOrInsertPosition);
+      auto convertedType =
+          VectorType::Builder(llvm::cast<VectorType>(sourceVector.getType()))
+              .setDim(0, resultType.getShape().back());
+      auto convertedVector =
+          rewriter.create<IntrOp>(loc, TypeRange{convertedType}, sourceVector);
+      result = rewriter.create<vector::InsertOp>(loc, convertedVector, result,
+                                                 extractOrInsertPosition);
+    }
+
+    rewriter.replaceOp(convertOp, result);
+    return success();
+  }
+};
+
+using ConvertToSvboolOpLowering =
+    SvboolConversionOpLowering<ConvertToSvboolOp, ConvertToSvboolIntrOp>;
+
+using ConvertFromSvboolOpLowering =
+    SvboolConversionOpLowering<ConvertFromSvboolOp, ConvertFromSvboolIntrOp>;
+
+} // namespace
+
 /// Populate the given list with patterns that convert from ArmSVE to LLVM.
 void mlir::populateArmSVELegalizeForLLVMExportPatterns(
     LLVMTypeConverter &converter, RewritePatternSet &patterns) {
@@ -88,7 +161,9 @@ void mlir::populateArmSVELegalizeForLLVMExportPatterns(
                ScalableMaskedMulFOpLowering,
                ScalableMaskedSDivIOpLowering,
                ScalableMaskedUDivIOpLowering,
-               ScalableMaskedDivFOpLowering>(converter);
+               ScalableMaskedDivFOpLowering,
+               ConvertToSvboolOpLowering,
+               ConvertFromSvboolOpLowering>(converter);
   // clang-format on
 }
 
@@ -107,7 +182,9 @@ void mlir::configureArmSVELegalizeForExportTarget(
                     ScalableMaskedMulFIntrOp,
                     ScalableMaskedSDivIIntrOp,
                     ScalableMaskedUDivIIntrOp,
-                    ScalableMaskedDivFIntrOp>();
+                    ScalableMaskedDivFIntrOp,
+                    ConvertToSvboolIntrOp,
+                    ConvertFromSvboolIntrOp>();
   target.addIllegalOp<SdotOp,
                       SmmlaOp,
                       UdotOp,
@@ -120,6 +197,8 @@ void mlir::configureArmSVELegalizeForExportTarget(
                       ScalableMaskedMulFOp,
                       ScalableMaskedSDivIOp,
                       ScalableMaskedUDivIOp,
-                      ScalableMaskedDivFOp>();
+                      ScalableMaskedDivFOp,
+                      ConvertToSvboolOp,
+                      ConvertFromSvboolOp>();
   // clang-format on
 }

mlir/test/Dialect/ArmSVE/invalid.mlir

@@ -0,0 +1,51 @@
+// RUN: mlir-opt %s -split-input-file -verify-diagnostics
+
+// -----
+
+func.func @arm_sve_convert_from_svbool__bad_mask_type(%bool: vector<2x[16]xi1>) -> vector<2x[8]xi2> {
+  // expected-error@+1 {{'result' must be trailing scalable vector of 1-bit signless integer values with dim -1 having a size of {16, 8, 4, 2, 1}, but got 'vector<2x[8]xi2>'}}
+  %mask = arm_sve.convert_from_svbool %bool : vector<2x[8]xi2>
+  return %mask : vector<2x[8]xi2>
+}
+
+// -----
+
+func.func @arm_sve_convert_from_svbool__bad_mask_shape(%bool : vector<[16]xi1>) -> vector<[7]xi1> {
+  // expected-error@+1 {{'result' must be trailing scalable vector of 1-bit signless integer values with dim -1 having a size of {16, 8, 4, 2, 1}, but got 'vector<[7]xi1>'}}
+  %mask = arm_sve.convert_from_svbool %bool : vector<[7]xi1>
+  return %mask : vector<[7]xi1>
+}
+
+// -----
+
+func.func @arm_sve_convert_from_svbool__bad_mask_scalability(%bool : vector<[4]x[16]xi1>) -> vector<[4]x[8]xi1> {
+  // expected-error@+1 {{'result' must be trailing scalable vector of 1-bit signless integer values with dim -1 having a size of {16, 8, 4, 2, 1}, but got 'vector<[4]x[8]xi1>'}}
+  %mask = arm_sve.convert_from_svbool %bool : vector<[4]x[8]xi1>
+  return %mask : vector<[4]x[8]xi1>
+}
+
+// -----
+
+func.func @arm_sve_convert_to_svbool__bad_mask_type(%mask: vector<2x[8]xi2>) -> vector<2x[16]xi1> {
+  // expected-error@+1 {{'source' must be trailing scalable vector of 1-bit signless integer values with dim -1 having a size of {16, 8, 4, 2, 1}, but got 'vector<2x[8]xi2>'}}
+  %bool = arm_sve.convert_to_svbool %mask : vector<2x[8]xi2>
+  return %bool : vector<2x[16]xi1>
+}
+
+// -----
+
+func.func @arm_sve_convert_to_svbool__bad_mask_shape(%mask : vector<[7]xi1>) -> vector<[16]xi1> {
+  // expected-error@+1 {{'source' must be trailing scalable vector of 1-bit signless integer values with dim -1 having a size of {16, 8, 4, 2, 1}, but got 'vector<[7]xi1>'}}
+  %bool = arm_sve.convert_to_svbool %mask : vector<[7]xi1>
+  return
+}
+
+// -----
+
+func.func @arm_sve_convert_to_svbool__bad_mask_scalability(%mask : vector<[4]x[8]xi1>) -> vector<[4]x[16]xi1> {
+  // expected-error@+1 {{'source' must be trailing scalable vector of 1-bit signless integer values with dim -1 having a size of {16, 8, 4, 2, 1}, but got 'vector<[4]x[8]xi1>'}}
+  %bool = arm_sve.convert_to_svbool %mask : vector<[4]x[8]xi1>
+  return
+}
+
+