[Matrix] Add support for matrix-by-scalar division.

[fhahn]

Summary

This patch extends the matrix spec to allow matrix-by-scalar division.

Originally support for / was left out to avoid ambiguity for the
matrix-matrix version of /, which could either be elementwise or
specified as matrix multiplication M1 * (1/M2).

For the matrix-scalar version, no ambiguity exists; * is also
an elementwise operation in that case. Matrix-by-scalar division
is commonly supported by systems including Matlab, Mathematica
or NumPy.

[everton.constantino]

clang/test/CodeGen/matrix-type-operators.c

717–827

Shouldn't a test for half floating point be added here as well?

[fhahn]

rebase & added a matrix_types_scalar_division feature, which allows users to check if the current version of clang is new enough to support it.

[fhahn]

clang/test/CodeGen/matrix-type-operators.c

717–827

Thanks for taking a look! I don't think we have any tests for half floating points at the moment. The tests do not cover each possible combination of types I think, because they re-use the existing code-gen for the stuff that's element type specific (except deciding whether to use floating point or integer ops).

I think it would probably be good to have some FP16 tests though, but that should be a separate change IMO.

[Gerolf]

Nice! This is a straightforward extension with minor code clean-up.

[everton.constantino]

clang/test/CodeGen/matrix-type-operators.c

717–827

Thanks for the reply. Ok, that makes sense, I was under the false impression FP16 was already covered.

[rjmccall]

clang/docs/MatrixTypes.rst

127

You don't need to be so formal in this paragraph, because you're about to get into a more formal treatment. I'd just say something like:

Given two matrixes, the `+ and - operators perform element-wise addition and subtraction, while the * operator performs matrix multiplication. +, -, *, and /` can also be used with a matrix and a scalar value, applying the operation to each element of the matrix.

Earlier versions of this extension did not support division by a scalar. You can test for the availability of this feature with `__has_extension(matrix_types_scalar_division)`.

131

You need to rework this paragraph, something like:

For the expression `M1 BIN_OP M2` where

• `BIN_OP is one of + or -, one of M1 and M2` is of matrix type, and the other is of matrix type or real type; or
• `BIN_OP is *, one of M1 and M2` is of matrix type, and the other is of a real type; or
• `BIN_OP is /, M1 is of matrix type, and M2` is of a real type:

clang/include/clang/Basic/LangOptions.def

399

Why is this a language option?

[fhahn]

Thank you very much John! I applied your suggestions to the wording and removed the unneeded language feature.

[rjmccall]

LGTM

clang/docs/MatrixTypes.rst

 
 Matrix Type Binary Operators
 ----------------------------
 
 Each matrix type supports the following binary operators: ``+``, ``-`` and ``*``. The ``*``
 operator provides matrix multiplication, while ``+`` and ``-`` are performed
 element-wise. There are also scalar versions of the operators, which take a
 matrix type and the matrix element type. The operation is applied to all
-elements of the matrix using the scalar value.
+elements of the matrix using the scalar value. In addition, the ``/`` operator
+is only supported for expression ``M1 / S1``, where ``M1`` is of matrix type
+and ``S1`` is of a real type.
 
-For ``BIN_OP`` in ``+``, ``-``, ``*`` given the expression ``M1 BIN_OP M2`` where
+For ``BIN_OP`` in ``+``, ``-``, ``*``, ``/`` given the expression ``M1 BIN_OP M2`` where
 at least one of ``M1`` or ``M2`` is of matrix type and, for `*`, the other is of
 a real type:
 
 * The usual arithmetic conversions are applied to ``M1`` and ``M2``. [ Note: if ``M1`` or
   ``M2`` are of a real type, they are broadcast to matrices here. — end note ]
 * ``M1`` and ``M2`` shall be of the same matrix type.
 * The result is equivalent to Res in the following where col is the number of
   columns and row is the number of rows in the matrix type:
 
 .. code-block:: c++

clang/include/clang/Basic/Features.def

 // Miscellaneous language extensions
 EXTENSION(overloadable_unmarked, true)
 EXTENSION(pragma_clang_attribute_namespaces, true)
 EXTENSION(pragma_clang_attribute_external_declaration, true)
 EXTENSION(statement_attributes_with_gnu_syntax, true)
 EXTENSION(gnu_asm, LangOpts.GNUAsm)
 EXTENSION(gnu_asm_goto_with_outputs, LangOpts.GNUAsm)
 EXTENSION(matrix_types, LangOpts.MatrixTypes)
+EXTENSION(matrix_types_scalar_division, LangOpts.MatrixTypesScalarDiv)
 
 FEATURE(cxx_abi_relative_vtable, LangOpts.CPlusPlus && LangOpts.RelativeCXXABIVTables)
 
 #undef EXTENSION
 #undef FEATURE

clang/include/clang/Basic/LangOptions.def

 
 LANGOPT(FixedPoint, 1, 0, "fixed point types")
 LANGOPT(PaddingOnUnsignedFixedPoint, 1, 0,
         "unsigned fixed point types having one extra padding bit")
 
 LANGOPT(RegisterStaticDestructors, 1, 1, "Register C++ static destructors")
 
 LANGOPT(MatrixTypes, 1, 0, "Enable or disable the builtin matrix type")
+LANGOPT(MatrixTypesScalarDiv, 1, 1, "Enable or disable scalar division support "
+                                    "for matrix types")
 
 COMPATIBLE_VALUE_LANGOPT(MaxTokens, 32, 0, "Max number of tokens per TU or 0")
 
 ENUM_LANGOPT(SignReturnAddressScope, SignReturnAddressScopeKind, 2, SignReturnAddressScopeKind::None,
              "Scope of return address signing")
 ENUM_LANGOPT(SignReturnAddressKey, SignReturnAddressKeyKind, 1, SignReturnAddressKeyKind::AKey,
              "Key used for return address signing")
 LANGOPT(BranchTargetEnforcement, 1, 0, "Branch-target enforcement enabled")

clang/lib/AST/Type.cpp

if (const ExtIntType *IT = dyn_cast<ExtIntType>(CanonicalType))
     return IT->isUnsigned();
 
   return false;
 }
 
 bool Type::hasUnsignedIntegerRepresentation() const {
   if (const auto *VT = dyn_cast<VectorType>(CanonicalType))
     return VT->getElementType()->isUnsignedIntegerOrEnumerationType();
-  else
-    return isUnsignedIntegerOrEnumerationType();
+  if (const auto *VT = dyn_cast<MatrixType>(CanonicalType))
+    return VT->getElementType()->isUnsignedIntegerOrEnumerationType();
 }
 
 bool Type::isFloatingType() const {
   if (const auto *BT = dyn_cast<BuiltinType>(CanonicalType))
     return BT->getKind() >= BuiltinType::Half &&
            BT->getKind() <= BuiltinType::Float128;
   if (const auto *CT = dyn_cast<ComplexType>(CanonicalType))
     return CT->getElementType()->isFloatingType();

clang/lib/CodeGen/CGExprScalar.cpp

if ((CGF.SanOpts.has(SanitizerKind::IntegerDivideByZero) ||
                Ops.mayHaveFloatDivisionByZero()) {
       llvm::Value *Zero = llvm::Constant::getNullValue(ConvertType(Ops.Ty));
       llvm::Value *NonZero = Builder.CreateFCmpUNE(Ops.RHS, Zero);
       EmitBinOpCheck(std::make_pair(NonZero, SanitizerKind::FloatDivideByZero),
                      Ops);
     }
   }
 
+  if (Ops.Ty->isConstantMatrixType()) {
+    llvm::MatrixBuilder<CGBuilderTy> MB(Builder);
+    // We need to check the types of the operands of the operator to get the
+    // correct matrix dimensions.
+    auto *BO = cast<BinaryOperator>(Ops.E);
+    assert(
+        isa<ConstantMatrixType>(BO->getLHS()->getType().getCanonicalType()) &&
+        "first operand must be a matrix");
+    assert(BO->getRHS()->getType().getCanonicalType()->isArithmeticType() &&
+           "second operand must be an arithmetic type");
+    return MB.CreateScalarDiv(Ops.LHS, Ops.RHS,
+                              Ops.Ty->hasUnsignedIntegerRepresentation());
+  }
+
   if (Ops.LHS->getType()->isFPOrFPVectorTy()) {
     llvm::Value *Val;
     CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, Ops.FPFeatures);
     Val = Builder.CreateFDiv(Ops.LHS, Ops.RHS, "div");
     if (CGF.getLangOpts().OpenCL &&
         !CGF.CGM.getCodeGenOpts().CorrectlyRoundedDivSqrt) {
       // OpenCL v1.1 s7.4: minimum accuracy of single precision / is 2.5ulp
       // OpenCL v1.2 s5.6.4.2: The -cl-fp32-correctly-rounded-divide-sqrt

clang/lib/Sema/SemaExpr.cpp

S.DiagRuntimeBehavior(Loc, RHS.get(),
                             << IsDiv << RHS.get()->getSourceRange());
 }
 
 QualType Sema::CheckMultiplyDivideOperands(ExprResult &LHS, ExprResult &RHS,
                                            SourceLocation Loc,
                                            bool IsCompAssign, bool IsDiv) {
   checkArithmeticNull(*this, LHS, RHS, Loc, /*IsCompare=*/false);
 
-  if (LHS.get()->getType()->isVectorType() ||
-      RHS.get()->getType()->isVectorType())
+  QualType LHSTy = LHS.get()->getType();
+  QualType RHSTy = RHS.get()->getType();
+  if (LHSTy->isVectorType() || RHSTy->isVectorType())
     return CheckVectorOperands(LHS, RHS, Loc, IsCompAssign,
                                /*AllowBothBool*/getLangOpts().AltiVec,
                                /*AllowBoolConversions*/false);
-  if (!IsDiv && (LHS.get()->getType()->isConstantMatrixType() ||
-                 RHS.get()->getType()->isConstantMatrixType()))
+  if (!IsDiv &&
+      (LHSTy->isConstantMatrixType() || RHSTy->isConstantMatrixType()))
     return CheckMatrixMultiplyOperands(LHS, RHS, Loc, IsCompAssign);
+  // For division, only matrix-by-scalar is supported. Other combinations with
+  // matrix types are invalid.
+  if (IsDiv && LHSTy->isConstantMatrixType() && RHSTy->isArithmeticType() &&
+      getLangOpts().MatrixTypesScalarDiv)
+    return CheckMatrixElementwiseOperands(LHS, RHS, Loc, IsCompAssign);
 
   QualType compType = UsualArithmeticConversions(
       LHS, RHS, Loc, IsCompAssign ? ACK_CompAssign : ACK_Arithmetic);
   if (LHS.isInvalid() || RHS.isInvalid())
     return QualType();
 
 
   if (compType.isNull() || !compType->isArithmeticType())

clang/test/CodeGen/matrix-type-operators.c

 // CHECK-NEXT:    [[MAT:%.*]] = load <27 x i32>, <27 x i32>* [[MAT_ADDR]], align 4
 // CHECK-NEXT:    [[RES:%.*]] = mul <27 x i32> <i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5>, [[MAT]]
 // CHECK-NEXT:    store <27 x i32> [[RES]], <27 x i32>* [[MAT_ADDR]], align 4
 // CHECK-NEXT:    ret void
 //
 void multiply_int_matrix_constant(ix9x3_t a) {
   a = 5 * a;
 }
 
 // CHECK-LABEL: @multiply_compound_int_matrix_constant(
 // CHECK-NEXT:  entry:
 // CHECK-NEXT:    [[A_ADDR:%.*]] = alloca [27 x i32], align 4
 // CHECK-NEXT:    [[MAT_ADDR:%.*]] = bitcast [27 x i32]* [[A_ADDR]] to <27 x i32>*
 // CHECK-NEXT:    store <27 x i32> [[A:%.*]], <27 x i32>* [[MAT_ADDR]], align 4
 // CHECK-NEXT:    [[MAT:%.*]] = load <27 x i32>, <27 x i32>* [[MAT_ADDR]], align 4
 // CHECK-NEXT:    [[RES:%.*]] = mul <27 x i32> [[MAT]], <i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5>
 // CHECK-NEXT:    store <27 x i32> [[RES]], <27 x i32>* [[MAT_ADDR]], align 4
 // CHECK-NEXT:    ret void
 //
 void multiply_compound_int_matrix_constant(ix9x3_t a) {
   a *= 5;
 }
 
+// CHECK-LABEL: @divide_double_matrix_scalar_float(
+// CHECK:         [[A:%.*]] = load <25 x double>, <25 x double>* {{.*}}, align 8
+// CHECK-NEXT:    [[S:%.*]] = load float, float* %s.addr, align 4
+// CHECK-NEXT:    [[S_EXT:%.*]] = fpext float [[S]] to double
+// CHECK-NEXT:    [[VECINSERT:%.*]] = insertelement <25 x double> poison, double [[S_EXT]], i32 0
+// CHECK-NEXT:    [[VECSPLAT:%.*]] = shufflevector <25 x double> [[VECINSERT]], <25 x double> poison, <25 x i32> zeroinitializer
+// CHECK-NEXT:    [[RES:%.*]] = fdiv <25 x double> [[A]], [[VECSPLAT]]
+// CHECK-NEXT:    store <25 x double> [[RES]], <25 x double>* {{.*}}, align 8
+// CHECK-NEXT:    ret void
+//
+void divide_double_matrix_scalar_float(dx5x5_t a, float s) {
+  a = a / s;
+}
+
+// CHECK-LABEL: @divide_double_matrix_scalar_double(
+// CHECK:         [[A:%.*]] = load <25 x double>, <25 x double>* {{.*}}, align 8
+// CHECK-NEXT:    [[S:%.*]] = load double, double* %s.addr, align 8
+// CHECK-NEXT:    [[VECINSERT:%.*]] = insertelement <25 x double> poison, double [[S]], i32 0
+// CHECK-NEXT:    [[VECSPLAT:%.*]] = shufflevector <25 x double> [[VECINSERT]], <25 x double> poison, <25 x i32> zeroinitializer
+// CHECK-NEXT:    [[RES:%.*]] = fdiv <25 x double> [[A]], [[VECSPLAT]]
+// CHECK-NEXT:    store <25 x double> [[RES]], <25 x double>* {{.*}}, align 8
+// CHECK-NEXT:    ret void
+//
+void divide_double_matrix_scalar_double(dx5x5_t a, double s) {
+  a = a / s;
+}
+
+// CHECK-LABEL: @divide_float_matrix_scalar_double(
+// CHECK:         [[MAT:%.*]] = load <6 x float>, <6 x float>* [[MAT_ADDR:%.*]], align 4
+// CHECK-NEXT:    [[S:%.*]] = load double, double* %s.addr, align 8
+// CHECK-NEXT:    [[S_TRUNC:%.*]] = fptrunc double [[S]] to float
+// CHECK-NEXT:    [[VECINSERT:%.*]] = insertelement <6 x float> poison, float [[S_TRUNC]], i32 0
+// CHECK-NEXT:    [[VECSPLAT:%.*]] = shufflevector <6 x float> [[VECINSERT]], <6 x float> poison, <6 x i32> zeroinitializer
+// CHECK-NEXT:    [[RES:%.*]] = fdiv <6 x float> [[MAT]], [[VECSPLAT]]
+// CHECK-NEXT:    store <6 x float> [[RES]], <6 x float>* [[MAT_ADDR]], align 4
+// CHECK-NEXT:    ret void
+//
+void divide_float_matrix_scalar_double(fx2x3_t b, double s) {
+  b = b / s;
+}
+
+// CHECK-LABEL: @divide_int_matrix_scalar_short(
+// CHECK:         [[MAT:%.*]] = load <27 x i32>, <27 x i32>* [[MAT_ADDR:%.*]], align 4
+// CHECK-NEXT:    [[S:%.*]] = load i16, i16* %s.addr, align 2
+// CHECK-NEXT:    [[S_EXT:%.*]] = sext i16 [[S]] to i32
+// CHECK-NEXT:    [[VECINSERT:%.*]] = insertelement <27 x i32> poison, i32 [[S_EXT]], i32 0
+// CHECK-NEXT:    [[VECSPLAT:%.*]] = shufflevector <27 x i32> [[VECINSERT]], <27 x i32> poison, <27 x i32> zeroinitializer
+// CHECK-NEXT:    [[RES:%.*]] = sdiv <27 x i32> [[MAT]], [[VECSPLAT]]
+// CHECK-NEXT:    store <27 x i32> [[RES]], <27 x i32>* [[MAT_ADDR]], align 4
+// CHECK-NEXT:    ret void
+//
+void divide_int_matrix_scalar_short(ix9x3_t b, short s) {
+  b = b / s;
+}
+
+// CHECK-LABEL: @divide_int_matrix_scalar_ull(
+// CHECK:         [[MAT:%.*]] = load <27 x i32>, <27 x i32>* [[MAT_ADDR:%.*]], align 4
+// CHECK-NEXT:    [[S:%.*]] = load i64, i64* %s.addr, align 8
+// CHECK-NEXT:    [[S_TRUNC:%.*]] = trunc i64 [[S]] to i32
+// CHECK-NEXT:    [[VECINSERT:%.*]] = insertelement <27 x i32> poison, i32 [[S_TRUNC]], i32 0
+// CHECK-NEXT:    [[VECSPLAT:%.*]] = shufflevector <27 x i32> [[VECINSERT]], <27 x i32> poison, <27 x i32> zeroinitializer
+// CHECK-NEXT:    [[RES:%.*]] = sdiv <27 x i32> [[MAT]], [[VECSPLAT]]
+// CHECK-NEXT:    store <27 x i32> [[RES]], <27 x i32>* [[MAT_ADDR]], align 4
+// CHECK-NEXT:    ret void
+//
+void divide_int_matrix_scalar_ull(ix9x3_t b, unsigned long long s) {
+  b = b / s;
+}
+
+// CHECK-LABEL: @divide_ull_matrix_scalar_ull(
+// CHECK:         [[MAT:%.*]] = load <8 x i64>, <8 x i64>* [[MAT_ADDR:%.*]], align 8
+// CHECK-NEXT:    [[S:%.*]] = load i64, i64* %s.addr, align 8
+// CHECK-NEXT:    [[VECINSERT:%.*]] = insertelement <8 x i64> poison, i64 [[S]], i32 0
+// CHECK-NEXT:    [[VECSPLAT:%.*]] = shufflevector <8 x i64> [[VECINSERT]], <8 x i64> poison, <8 x i32> zeroinitializer
+// CHECK-NEXT:    [[RES:%.*]] = udiv <8 x i64> [[MAT]], [[VECSPLAT]]
+// CHECK-NEXT:    store <8 x i64> [[RES]], <8 x i64>* [[MAT_ADDR]], align 8
+// CHECK-NEXT:    ret void
+//
+void divide_ull_matrix_scalar_ull(ullx4x2_t b, unsigned long long s) {
+  b = b / s;
+}
+
+// CHECK-LABEL: @divide_float_matrix_constant(
+// CHECK-NEXT:  entry:
+// CHECK-NEXT:    [[A_ADDR:%.*]] = alloca [6 x float], align 4
+// CHECK-NEXT:    [[MAT_ADDR:%.*]] = bitcast [6 x float]* [[A_ADDR]] to <6 x float>*
+// CHECK-NEXT:    store <6 x float> [[A:%.*]], <6 x float>* [[MAT_ADDR]], align 4
+// CHECK-NEXT:    [[MAT:%.*]] = load <6 x float>, <6 x float>* [[MAT_ADDR]], align 4
+// CHECK-NEXT:    [[RES:%.*]] = fdiv <6 x float> [[MAT]], <float 2.500000e+00, float 2.500000e+00, float 2.500000e+00, float 2.500000e+00, float 2.500000e+00, float 2.500000e+00>
+// CHECK-NEXT:    store <6 x float> [[RES]], <6 x float>* [[MAT_ADDR]], align 4
+// CHECK-NEXT:    ret void
+//
+void divide_float_matrix_constant(fx2x3_t a) {
+  a = a / 2.5;
+}
+
 // Tests for the matrix type operators.
 
 typedef double dx5x5_t __attribute__((matrix_type(5, 5)));
 typedef float fx2x3_t __attribute__((matrix_type(2, 3)));
 
 // Check that we can use matrix index expression on different floating point
 // matrixes and indices.
 void insert_double_matrix_const_idx_ll_u_double(dx5x5_t a, double d, fx2x3_t b, float e, int j, unsigned k) {

clang/test/CodeGen/matrix-type.c

 // RUN: %clang_cc1 -fenable-matrix -triple x86_64-apple-darwin %s -emit-llvm -disable-llvm-passes -o - | FileCheck %s
 
 #if !__has_extension(matrix_types)
 #error Expected extension 'matrix_types' to be enabled
 #endif
 
+#if !__has_extension(matrix_types_scalar_division)
+#error Expected extension 'matrix_types_scalar_division' to be enabled
+#endif
+
 typedef double dx5x5_t __attribute__((matrix_type(5, 5)));
 
 // CHECK: %struct.Matrix = type { i8, [12 x float], float }
 
 void load_store_double(dx5x5_t *a, dx5x5_t *b) {
   // CHECK-LABEL:  define{{.*}} void @load_store_double(
   // CHECK-NEXT:  entry:
   // CHECK-NEXT:    %a.addr = alloca [25 x double]*, align 8

clang/test/Sema/matrix-type-operators.c

void mat_scalar_multiply(sx10x10_t a, sx5x10_t b, float sf, char *p) {
   p *= a;
   // expected-error@-1 {{casting 'char *' to incompatible type 'float'}}
   // expected-error@-2 {{invalid operands to binary expression ('char *' and 'sx10x10_t' (aka 'float __attribute__((matrix_type(10, 10)))'))}}
 
   sf = a * sf;
   // expected-error@-1 {{assigning to 'float' from incompatible type 'sx10x10_t' (aka 'float __attribute__((matrix_type(10, 10)))')}}
 }
 
+void mat_scalar_divide(sx10x10_t a, sx5x10_t b, float sf, char *p) {
+  // Shape of multiplication result does not match the type of b.
+  b = a / sf;
+  // expected-error@-1 {{assigning to 'sx5x10_t' (aka 'float __attribute__((matrix_type(5, 10)))') from incompatible type 'sx10x10_t' (aka 'float __attribute__((matrix_type(10, 10)))')}}
+  b = sf / a;
+  // expected-error@-1 {{invalid operands to binary expression ('float' and 'sx10x10_t' (aka 'float __attribute__((matrix_type(10, 10)))'))}}
+
+  a = a / p;
+  // expected-error@-1 {{invalid operands to binary expression ('sx10x10_t' (aka 'float __attribute__((matrix_type(10, 10)))') and 'char *')}}
+  a = p / a;
+  // expected-error@-1 {{invalid operands to binary expression ('char *' and 'sx10x10_t' (aka 'float __attribute__((matrix_type(10, 10)))'))}}
+
+  sf = a / sf;
+  // expected-error@-1 {{assigning to 'float' from incompatible type 'sx10x10_t' (aka 'float __attribute__((matrix_type(10, 10)))')}}
+}
+
+void matrix_matrix_divide(sx10x10_t a, sx5x10_t b, ix10x5_t c, ix10x10_t d, float sf, char *p) {
+  // Matrix by matrix division is not supported.
+  a = a / a;
+  // expected-error@-1 {{invalid operands to binary expression ('sx10x10_t' (aka 'float __attribute__((matrix_type(10, 10)))') and 'sx10x10_t')}}
+
+  b = a / a;
+  // expected-error@-1 {{invalid operands to binary expression ('sx10x10_t' (aka 'float __attribute__((matrix_type(10, 10)))') and 'sx10x10_t')}}
+
+  // Check element type mismatches.
+  a = b / c;
+  // expected-error@-1 {{invalid operands to binary expression ('sx5x10_t' (aka 'float __attribute__((matrix_type(5, 10)))') and 'ix10x5_t' (aka 'int __attribute__((matrix_type(10, 5)))'))}}
+  d = a / a;
+  // expected-error@-1 {{invalid operands to binary expression ('sx10x10_t' (aka 'float __attribute__((matrix_type(10, 10)))') and 'sx10x10_t')}}
+
+  p = a / a;
+  // expected-error@-1 {{invalid operands to binary expression ('sx10x10_t' (aka 'float __attribute__((matrix_type(10, 10)))') and 'sx10x10_t')}}
+}
+
 sx5x10_t get_matrix();
 
 void insert(sx5x10_t a, float f) {
   // Non integer indexes.
   a[3][f] = 0;
   // expected-error@-1 {{matrix column index is not an integer}}
   a[f][9] = 0;
   // expected-error@-1 {{matrix row index is not an integer}}

llvm/include/llvm/IR/MatrixBuilder.h

public:
   /// RHS.
   Value *CreateScalarMultiply(Value *LHS, Value *RHS) {
     std::tie(LHS, RHS) = splatScalarOperandIfNeeded(LHS, RHS);
     if (LHS->getType()->getScalarType()->isFloatingPointTy())
       return B.CreateFMul(LHS, RHS);
     return B.CreateMul(LHS, RHS);
   }
 
+  /// Divide matrix \p LHS by scalar \p RHS. If the operands are integers, \p
+  /// IsUnsigned indicates whether UDiv or SDiv should be used.
+  Value *CreateScalarDiv(Value *LHS, Value *RHS, bool IsUnsigned) {
+    assert(LHS->getType()->isVectorTy() && !RHS->getType()->isVectorTy());
+    assert(!isa<ScalableVectorType>(LHS->getType()) &&
+           "LHS Assumed to be fixed width");
+    RHS =
+        B.CreateVectorSplat(cast<VectorType>(LHS->getType())->getElementCount(),
+                            RHS, "scalar.splat");
+    return cast<VectorType>(LHS->getType())
+                   ->getElementType()
+                   ->isFloatingPointTy()
+               ? B.CreateFDiv(LHS, RHS)
+               : (IsUnsigned ? B.CreateUDiv(LHS, RHS) : B.CreateSDiv(LHS, RHS));
+  }
+
   /// Extracts the element at (\p RowIdx, \p ColumnIdx) from \p Matrix.
   Value *CreateExtractElement(Value *Matrix, Value *RowIdx, Value *ColumnIdx,
                               unsigned NumRows, Twine const &Name = "") {
 
     unsigned MaxWidth = std::max(RowIdx->getType()->getScalarSizeInBits(),
                                  ColumnIdx->getType()->getScalarSizeInBits());
     Type *IntTy = IntegerType::get(RowIdx->getType()->getContext(), MaxWidth);
     RowIdx = B.CreateZExt(RowIdx, IntTy);