[mlir][math] Initial support for fastmath flag attributes for Math dialect.

[vzakhari]

Summary

Added arith::FastMathAttr and ArithFastMathInterface support for Math dialect
floating point operations.

This change-set creates ArithCommon conversion utils that currently
provide classes and methods to aid with arith::FastMathAttr conversion
into LLVM::FastmathFlags. These utils are used in ArithToLLVM and
MathToLLVM convertors, but may eventually be used by other converters
that need to convert fast math attributes.

Since Math dialect operations use arith::FastMathAttr, MathOps.td now
has to include enum and attributes definitions from Arith dialect.
To minimize the amount of TD code included from Arith dialect,
I moved FastMathAttr definition into ArithBase.td.

Depends on D136225

[vzakhari]

Rebase on top of the latest D126305.

[kiranchandramohan]

LGTM. Please wait a day before you submit in case there are comments from other reviewers.

mlir/lib/Conversion/ArithCommon/CMakeLists.txt

1 ↗

(On Diff #472400)

Nit: Would it be better to retain AttrToLLVM in the name of the pass like MLIRArithAttrToLLVMConversion? Or do you anticipate other conversions to be present here?

[vzakhari]

mlir/lib/Conversion/ArithCommon/CMakeLists.txt

1 ↗

(On Diff #472400)

Yes, I created ArithCommon directory with the intention that it would hold some common conversion utils from Arith dialect entities (like FastMathFlagsAttr) to other dialects. For example, there are ArithToSPIRV and MathToSPIRV convertors, which may also have some common utils hosted here. I am not sure though if these SPIRV utils will necessarily be put into the same CMake component (e.g. the SPIRV convertors may not need a dependency on LLVM dialect), so I guess renaming it to MLIRArithAttrToLLVMConversion should be appropriate right now.

mlir/include/mlir/Dialect/Math/IR/CMakeLists.txt

+set(LLVM_TARGET_DEFINITIONS MathOps.td)
+mlir_tablegen(MathOpsEnums.h.inc -gen-enum-decls)
+mlir_tablegen(MathOpsEnums.cpp.inc -gen-enum-defs)
+mlir_tablegen(MathOpsAttributes.h.inc -gen-attrdef-decls
+              -attrdefs-dialect=math)
+mlir_tablegen(MathOpsAttributes.cpp.inc -gen-attrdef-defs
+              -attrdefs-dialect=math)
+add_public_tablegen_target(MLIRMathOpsAttributesIncGen)
+
+set(LLVM_TARGET_DEFINITIONS MathOpsInterfaces.td)
+mlir_tablegen(MathOpsInterfaces.h.inc -gen-op-interface-decls)
+mlir_tablegen(MathOpsInterfaces.cpp.inc -gen-op-interface-defs)
+add_public_tablegen_target(MLIRMathOpsInterfacesIncGen)
+
 add_mlir_dialect(MathOps math)
 add_mlir_doc(MathOps MathOps Dialects/ -gen-dialect-doc)

mlir/include/mlir/Dialect/Math/IR/Math.h

 
 //===----------------------------------------------------------------------===//
 // Math Dialect
 //===----------------------------------------------------------------------===//
 
 #include "mlir/Dialect/Math/IR/MathOpsDialect.h.inc"
 
 //===----------------------------------------------------------------------===//
+// Math Dialect Enum Attributes
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Dialect/Math/IR/MathOpsEnums.h.inc"
+#define GET_ATTRDEF_CLASSES
+#include "mlir/Dialect/Math/IR/MathOpsAttributes.h.inc"
+
+//===----------------------------------------------------------------------===//
+// Math Interfaces
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Dialect/Math/IR/MathOpsInterfaces.h.inc"
+
+//===----------------------------------------------------------------------===//
 // Math Dialect Operations
 //===----------------------------------------------------------------------===//
 
 #define GET_OP_CLASSES
 #include "mlir/Dialect/Math/IR/MathOps.h.inc"
 
 #endif // MLIR_DIALECT_MATH_IR_MATH_H_

mlir/include/mlir/Dialect/Math/IR/MathBase.td

 //===- MathBase.td - Base definitions for math dialect ------*- tablegen -*-==//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 #ifndef MATH_BASE
 #define MATH_BASE
+include "mlir/IR/EnumAttr.td"
 include "mlir/IR/OpBase.td"
 def Math_Dialect : Dialect {
   let name = "math";
   let cppNamespace = "::mlir::math";
   let description = [{
     The math dialect is intended to hold mathematical operations on integer and
     floating types beyond simple arithmetics. Each operation works on scalar, vector
     or tensor type. On vector and tensor type operations apply elementwise unless
     explicitly specified otherwise. As an example, the floating point absolute value
     can be expressed as:
 
     ```mlir
     // Scalar absolute value.
     %a = math.absf %b : f64
 
     // Vector elementwise absolute value.
     %f = math.absf %g : vector<4xf32>
 
     // Tensor elementwise absolute value.
     %x = math.absf %y : tensor<4x?xf8>
     ```
   }];
   let hasConstantMaterializer = 1;
+  let useDefaultAttributePrinterParser = 1;
+}
+
+def FASTMATH_NONE            : I32BitEnumAttrCaseNone<"none"      >;
+def FASTMATH_REASSOC         : I32BitEnumAttrCaseBit<"reassoc",  0>;
+def FASTMATH_NO_NANS         : I32BitEnumAttrCaseBit<"nnan",     1>;
+def FASTMATH_NO_INFS         : I32BitEnumAttrCaseBit<"ninf",     2>;
+def FASTMATH_NO_SIGNED_ZEROS : I32BitEnumAttrCaseBit<"nsz",      3>;
+def FASTMATH_ALLOW_RECIP     : I32BitEnumAttrCaseBit<"arcp",     4>;
+def FASTMATH_ALLOW_CONTRACT  : I32BitEnumAttrCaseBit<"contract", 5>;
+def FASTMATH_APPROX_FUNC     : I32BitEnumAttrCaseBit<"afn",      6>;
+def FASTMATH_FAST            : I32BitEnumAttrCaseGroup<
+    "fast",
+    [
+      FASTMATH_REASSOC,         FASTMATH_NO_NANS,     FASTMATH_NO_INFS,
+      FASTMATH_NO_SIGNED_ZEROS, FASTMATH_ALLOW_RECIP, FASTMATH_ALLOW_CONTRACT,
+      FASTMATH_APPROX_FUNC]>;
+
+def FastMathFlags : I32BitEnumAttr<
+    "FastMathFlags",
+    "Floating point fast math flags",
+    [
+      FASTMATH_NONE,           FASTMATH_REASSOC,         FASTMATH_NO_NANS,
+      FASTMATH_NO_INFS,        FASTMATH_NO_SIGNED_ZEROS, FASTMATH_ALLOW_RECIP,
+      FASTMATH_ALLOW_CONTRACT, FASTMATH_APPROX_FUNC,     FASTMATH_FAST]> {
+  let separator = ",";
+  let cppNamespace = "::mlir::math";
+  let genSpecializedAttr = 0;
+  let printBitEnumPrimaryGroups = 1;
 }
 #endif // MATH_BASE

mlir/include/mlir/Dialect/Math/IR/MathOps.td

 //===- MathOps.td - Math op definitions --------------------*- tablegen -*-===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef MATH_OPS
 #define MATH_OPS
 
 include "mlir/Dialect/Math/IR/MathBase.td"
+include "mlir/Dialect/Math/IR/MathOpsInterfaces.td"
 include "mlir/Interfaces/InferTypeOpInterface.td"
 include "mlir/Interfaces/VectorInterfaces.td"
 include "mlir/Interfaces/SideEffectInterfaces.td"
 
+def Math_FastMathAttr :
+    EnumAttr<Math_Dialect, FastMathFlags, "fastmath"> {
+  let assemblyFormat = "`<` $value `>`";
+}
+
 // Base class for math dialect ops. Ops in this dialect have no side effects and
 // can be applied element-wise to vectors and tensors.
 class Math_Op<string mnemonic, list<Trait> traits = []> :
     Op<Math_Dialect, mnemonic, traits # [Pure,
     DeclareOpInterfaceMethods<VectorUnrollOpInterface>] #
     ElementwiseMappable.traits>;
 
 // Base class for unary math operations on integer types. Require an operand
 // and result of the same type. This type can be an integer type, vector or
 // tensor thereof.
 class Math_IntegerUnaryOp<string mnemonic, list<Trait> traits = []> :
     Math_Op<mnemonic, traits # [SameOperandsAndResultType]> {
   let arguments = (ins SignlessIntegerLike:$operand);
   let results = (outs SignlessIntegerLike:$result);
 
   let assemblyFormat = "$operand attr-dict `:` type($result)";
 }
 
 // Base class for unary math operations on floating point types. Require an
 // operand and result of the same type. This type can be a floating point type,
 // vector or tensor thereof.
 class Math_FloatUnaryOp<string mnemonic, list<Trait> traits = []> :
-    Math_Op<mnemonic, traits # [SameOperandsAndResultType]> {
-  let arguments = (ins FloatLike:$operand);
+    Math_Op<mnemonic,
+        traits # [SameOperandsAndResultType,
+                  DeclareOpInterfaceMethods<MathFastMathInterface>]> {
+  let arguments = (ins FloatLike:$operand,
+      DefaultValuedAttr<Math_FastMathAttr, "FastMathFlags::none">:$fastmath);
   let results = (outs FloatLike:$result);
 
-  let assemblyFormat = "$operand attr-dict `:` type($result)";
+  let assemblyFormat = "$operand custom<MathFastMathAttr>($fastmath) attr-dict `:` type($result)";
 }
 
 // Base class for binary math operations on integer types. Require two
 // operands and one result of the same type. This type can be an integer
 // type, vector or tensor thereof.
 class Math_IntegerBinaryOp<string mnemonic, list<Trait> traits = []> :
     Math_Op<mnemonic, traits # [SameOperandsAndResultType]> {
   let arguments = (ins SignlessIntegerLike:$lhs, SignlessIntegerLike:$rhs);
   let results = (outs SignlessIntegerLike:$result);
 
   let assemblyFormat = "$lhs `,` $rhs attr-dict `:` type($result)";
 }
 
 // Base class for binary math operations on floating point types. Require two
 // operands and one result of the same type. This type can be a floating point
 // type, vector or tensor thereof.
 class Math_FloatBinaryOp<string mnemonic, list<Trait> traits = []> :
-    Math_Op<mnemonic, traits # [SameOperandsAndResultType]> {
-  let arguments = (ins FloatLike:$lhs, FloatLike:$rhs);
+    Math_Op<mnemonic,
+        traits # [SameOperandsAndResultType,
+                  DeclareOpInterfaceMethods<MathFastMathInterface>]> {
+  let arguments = (ins FloatLike:$lhs, FloatLike:$rhs,
+      DefaultValuedAttr<Math_FastMathAttr, "FastMathFlags::none">:$fastmath);
   let results = (outs FloatLike:$result);
 
-  let assemblyFormat = "$lhs `,` $rhs attr-dict `:` type($result)";
+  let assemblyFormat = "$lhs `,` $rhs custom<MathFastMathAttr>($fastmath) attr-dict `:` type($result)";
 }
 
 // Base class for floating point ternary operations. Require three operands and
 // one result of the same type. This type can be a floating point type, vector
 // or tensor thereof.
 class Math_FloatTernaryOp<string mnemonic, list<Trait> traits = []> :
-    Math_Op<mnemonic, traits # [SameOperandsAndResultType]> {
-  let arguments = (ins FloatLike:$a, FloatLike:$b, FloatLike:$c);
+    Math_Op<mnemonic,
+        traits # [SameOperandsAndResultType,
+                  DeclareOpInterfaceMethods<MathFastMathInterface>]> {
+  let arguments = (ins FloatLike:$a, FloatLike:$b, FloatLike:$c,
+      DefaultValuedAttr<Math_FastMathAttr, "FastMathFlags::none">:$fastmath);
   let results = (outs FloatLike:$result);
 
-  let assemblyFormat = "$a `,` $b `,` $c attr-dict `:` type($result)";
+  let assemblyFormat = "$a `,` $b `,` $c custom<MathFastMathAttr>($fastmath) attr-dict `:` type($result)";
 }
 
 //===----------------------------------------------------------------------===//
 // AbsFOp
 //===----------------------------------------------------------------------===//
 
 def Math_AbsFOp : Math_FloatUnaryOp<"absf"> {
   let summary = "floating point absolute-value operation";

mlir/include/mlir/Dialect/Math/IR/MathOpsInterfaces.td

+//===-- MathOpsInterfaces.td - math op interfaces ----------*- tablegen -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This is the Math interfaces definition file.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef MATH_OPS_INTERFACES
+#define MATH_OPS_INTERFACES
+
+include "mlir/IR/OpBase.td"
+
+def MathFastMathInterface : OpInterface<"MathFastMathInterface"> {
+  let description = [{
+    Access to operation fastmath flags.
+  }];
+
+  let cppNamespace = "::mlir::math";
+
+  let methods = [
+    InterfaceMethod<
+      /*desc=*/        "Returns a FastMathFlagsAttr attribute for the operation",
+      /*returnType=*/  "FastMathFlagsAttr",
+      /*methodName=*/  "getFastMathFlagsAttr",
+      /*args=*/        (ins),
+      /*methodBody=*/  [{}],
+      /*defaultImpl=*/ [{
+        ConcreteOp op = cast<ConcreteOp>(this->getOperation());
+        return op.getFastmathAttr();
+      }]
+      >,
+    StaticInterfaceMethod<
+      /*desc=*/        [{Returns the name of the FastMathFlagsAttr attribute
+                         for the operation}],
+      /*returnType=*/  "StringRef",
+      /*methodName=*/  "getFastMathAttrName",
+      /*args=*/        (ins),
+      /*methodBody=*/  [{}],
+      /*defaultImpl=*/ [{
+        return "fastmath";
+      }]
+      >
+  ];
+}
+
+#endif // MATH_OPS_INTERFACES

mlir/lib/Conversion/LLVMCommon/CMakeLists.txt

 add_mlir_conversion_library(MLIRLLVMCommonConversion
   ConversionTarget.cpp
   LoweringOptions.cpp
   MemRefBuilder.cpp
   Pattern.cpp
   StructBuilder.cpp
   TypeConverter.cpp
   VectorPattern.cpp
 
   LINK_COMPONENTS
   Core
 
   LINK_LIBS PUBLIC
   MLIRIR
   MLIRLLVMDialect
+  MLIRMathDialect
   MLIRSupport
   MLIRTransforms
   MLIRArithDialect
   )

mlir/lib/Conversion/LLVMCommon/Pattern.cpp

 //===- Pattern.cpp - Conversion pattern to the LLVM dialect ---------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #include "mlir/Conversion/LLVMCommon/Pattern.h"
 #include "mlir/Dialect/Arith/IR/Arith.h"
 #include "mlir/Dialect/LLVMIR/FunctionCallUtils.h"
 #include "mlir/Dialect/LLVMIR/LLVMDialect.h"
 #include "mlir/Dialect/LLVMIR/LLVMTypes.h"
+#include "mlir/Dialect/Math/IR/Math.h"
 #include "mlir/IR/AffineMap.h"
 #include "mlir/IR/BuiltinAttributes.h"
 
 using namespace mlir;
 
 // Map arithmetic fastmath enum values to LLVMIR enum values.
 static LLVM::FastmathFlags
 convertArithFastMathFlagsToLLVM(arith::FastMathFlags arithFMF) {

 // Create an LLVM fastmath attribute from a given arithmetic fastmath attribute.
 static LLVM::FastmathFlagsAttr
 convertArithFastMathAttr(arith::FastMathFlagsAttr fmfAttr) {
   auto arithFMF = fmfAttr.getValue();
   return LLVM::FastmathFlagsAttr::get(
       fmfAttr.getContext(), convertArithFastMathFlagsToLLVM(arithFMF));
 }
 
+// Map Math dialect fastmath enum values to LLVMIR enum values.
+static LLVM::FastmathFlags
+convertMathFastMathFlagsToLLVM(math::FastMathFlags mathFMF) {
+  LLVM::FastmathFlags llvmFMF{};
+  const std::pair<math::FastMathFlags, LLVM::FastmathFlags> flags[] = {
+      {math::FastMathFlags::nnan, LLVM::FastmathFlags::nnan},
+      {math::FastMathFlags::ninf, LLVM::FastmathFlags::ninf},
+      {math::FastMathFlags::nsz, LLVM::FastmathFlags::nsz},
+      {math::FastMathFlags::arcp, LLVM::FastmathFlags::arcp},
+      {math::FastMathFlags::contract, LLVM::FastmathFlags::contract},
+      {math::FastMathFlags::afn, LLVM::FastmathFlags::afn},
+      {math::FastMathFlags::reassoc, LLVM::FastmathFlags::reassoc}};
+  for (auto fmfMap : flags) {
+    if (bitEnumContainsAny(mathFMF, fmfMap.first))
+      llvmFMF = llvmFMF | fmfMap.second;
+  }
+  return llvmFMF;
+}
+
+// Create an LLVM fastmath attribute from a given Math dialect fastmath
+// attribute.
+static LLVM::FastmathFlagsAttr
+convertMathFastMathAttr(math::FastMathFlagsAttr fmfAttr) {
+  auto mathFMF = fmfAttr.getValue();
+  return LLVM::FastmathFlagsAttr::get(fmfAttr.getContext(),
+                                      convertMathFastMathFlagsToLLVM(mathFMF));
+}
+
 static llvm::Optional<llvm::StringRef> getTargetFastmathAttr(MLIRContext *ctx,
                                                              StringRef target) {
   OperationName oName(target, ctx);
   llvm::Optional<llvm::StringRef> attrName;
   // Using the getInterface() method of OperationName here, even though that
   // method may be deprecated or removed.
   // TODO: Improve infrastructure for calling static interface methods
   if (auto fmfInterface = oName.getInterface<LLVM::FastmathFlagsInterface>())
auto arithFMFAttr = convertedAttrList.value()
                             .dyn_cast_or_null<arith::FastMathFlagsAttr>();
     if (arithFMFAttr) {
       llvm::Optional<llvm::StringRef> targetAttrName =
           getTargetFastmathAttr(srcOp->getContext(), targetOp);
       if (targetAttrName)
         convertedAttrList.value().set(targetAttrName.value(),
                                       convertArithFastMathAttr(arithFMFAttr));
     }
+  } else if (auto fmi = dyn_cast<math::MathFastMathInterface>(*srcOp)) {
+    // TODO: get rid of the code duplication.
+
+    // Get the name of the Math dialect fastmath attribute
+    StringRef mathFMFAttrName = fmi.getFastMathAttrName();
+    // Copy the source attributes
+    convertedAttrList = NamedAttrList{srcAttrs};
+    // Remove the source fastmath attribute. We will either replace it below
+    // (if the target supports the fastmath interface) or just "drop" it.
+    auto mathFMFAttr = convertedAttrList.value()
+                           .erase(mathFMFAttrName)
+                           .dyn_cast_or_null<math::FastMathFlagsAttr>();
+    if (mathFMFAttr) {
+      llvm::Optional<StringRef> targetAttrName =
+          getTargetFastmathAttr(srcOp->getContext(), targetOp);
+      if (targetAttrName)
+        convertedAttrList.value().set(targetAttrName.value(),
+                                      convertMathFastMathAttr(mathFMFAttr));
+    }
   }
 }

mlir/lib/Conversion/MathToLLVM/MathToLLVM.cpp

matchAndRewrite(math::ExpM1Op op, OpAdaptor adaptor,
 
     if (!operandType || !LLVM::isCompatibleType(operandType))
       return failure();
 
     auto loc = op.getLoc();
     auto resultType = op.getResult().getType();
     auto floatType = getElementTypeOrSelf(resultType).cast<FloatType>();
     auto floatOne = rewriter.getFloatAttr(floatType, 1.0);
+    LLVM::detail::ConvertToLLVMAttributes expAttrs(
+        op, LLVM::ExpOp::getOperationName());
+    LLVM::detail::ConvertToLLVMAttributes subAttrs(
+        op, LLVM::FSubOp::getOperationName());
 
     if (!operandType.isa<LLVM::LLVMArrayType>()) {
       LLVM::ConstantOp one;
       if (LLVM::isCompatibleVectorType(operandType)) {
         one = rewriter.create<LLVM::ConstantOp>(
             loc, operandType,
             SplatElementsAttr::get(resultType.cast<ShapedType>(), floatOne));
       } else {
         one = rewriter.create<LLVM::ConstantOp>(loc, operandType, floatOne);
       }
-      auto exp = rewriter.create<LLVM::ExpOp>(loc, adaptor.getOperand());
-      rewriter.replaceOpWithNewOp<LLVM::FSubOp>(op, operandType, exp, one);
+      auto exp = rewriter.create<LLVM::ExpOp>(loc, adaptor.getOperand(),
+                                              expAttrs.getAttrs());
+      rewriter.replaceOpWithNewOp<LLVM::FSubOp>(
+          op, operandType, ValueRange{exp, one}, subAttrs.getAttrs());
       return success();
     }
 
     auto vectorType = resultType.dyn_cast<VectorType>();
     if (!vectorType)
       return rewriter.notifyMatchFailure(op, "expected vector result type");
 
     return LLVM::detail::handleMultidimensionalVectors(
         op.getOperation(), adaptor.getOperands(), *getTypeConverter(),
         [&](Type llvm1DVectorTy, ValueRange operands) {
           auto splatAttr = SplatElementsAttr::get(
               mlir::VectorType::get(
                   {LLVM::getVectorNumElements(llvm1DVectorTy).getFixedValue()},
                   floatType),
               floatOne);
           auto one =
               rewriter.create<LLVM::ConstantOp>(loc, llvm1DVectorTy, splatAttr);
-          auto exp =
-              rewriter.create<LLVM::ExpOp>(loc, llvm1DVectorTy, operands[0]);
-          return rewriter.create<LLVM::FSubOp>(loc, llvm1DVectorTy, exp, one);
+          auto exp = rewriter.create<LLVM::ExpOp>(
+              loc, llvm1DVectorTy, operands[0], expAttrs.getAttrs());
+          return rewriter.create<LLVM::FSubOp>(
+              loc, llvm1DVectorTy, ValueRange{exp, one}, subAttrs.getAttrs());
         },
         rewriter);
   }
 };
 
 // A `log1p` is converted into `log(1 + ...)`.
 struct Log1pOpLowering : public ConvertOpToLLVMPattern<math::Log1pOp> {
   using ConvertOpToLLVMPattern<math::Log1pOp>::ConvertOpToLLVMPattern;
 
   LogicalResult
   matchAndRewrite(math::Log1pOp op, OpAdaptor adaptor,
                   ConversionPatternRewriter &rewriter) const override {
     auto operandType = adaptor.getOperand().getType();
 
     if (!operandType || !LLVM::isCompatibleType(operandType))
       return rewriter.notifyMatchFailure(op, "unsupported operand type");
 
     auto loc = op.getLoc();
     auto resultType = op.getResult().getType();
     auto floatType = getElementTypeOrSelf(resultType).cast<FloatType>();
     auto floatOne = rewriter.getFloatAttr(floatType, 1.0);
+    LLVM::detail::ConvertToLLVMAttributes addAttrs(
+        op, LLVM::FAddOp::getOperationName());
+    LLVM::detail::ConvertToLLVMAttributes logAttrs(
+        op, LLVM::LogOp::getOperationName());
 
     if (!operandType.isa<LLVM::LLVMArrayType>()) {
       LLVM::ConstantOp one =
           LLVM::isCompatibleVectorType(operandType)
               ? rewriter.create<LLVM::ConstantOp>(
                     loc, operandType,
                     SplatElementsAttr::get(resultType.cast<ShapedType>(),
                                            floatOne))
               : rewriter.create<LLVM::ConstantOp>(loc, operandType, floatOne);
 
-      auto add = rewriter.create<LLVM::FAddOp>(loc, operandType, one,
-                                               adaptor.getOperand());
-      rewriter.replaceOpWithNewOp<LLVM::LogOp>(op, operandType, add);
+      auto add = rewriter.create<LLVM::FAddOp>(
+          loc, operandType, ValueRange{one, adaptor.getOperand()},
+          addAttrs.getAttrs());
+      rewriter.replaceOpWithNewOp<LLVM::LogOp>(op, operandType, ValueRange{add},
+                                               logAttrs.getAttrs());
       return success();
     }
 
     auto vectorType = resultType.dyn_cast<VectorType>();
     if (!vectorType)
       return rewriter.notifyMatchFailure(op, "expected vector result type");
 
     return LLVM::detail::handleMultidimensionalVectors(
         op.getOperation(), adaptor.getOperands(), *getTypeConverter(),
         [&](Type llvm1DVectorTy, ValueRange operands) {
           auto splatAttr = SplatElementsAttr::get(
               mlir::VectorType::get(
                   {LLVM::getVectorNumElements(llvm1DVectorTy).getFixedValue()},
                   floatType),
               floatOne);
           auto one =
               rewriter.create<LLVM::ConstantOp>(loc, llvm1DVectorTy, splatAttr);
-          auto add = rewriter.create<LLVM::FAddOp>(loc, llvm1DVectorTy, one,
-                                                   operands[0]);
-          return rewriter.create<LLVM::LogOp>(loc, llvm1DVectorTy, add);
+          auto add = rewriter.create<LLVM::FAddOp>(loc, llvm1DVectorTy,
+                                                   ValueRange{one, operands[0]},
+                                                   addAttrs.getAttrs());
+          return rewriter.create<LLVM::LogOp>(
+              loc, llvm1DVectorTy, ValueRange{add}, logAttrs.getAttrs());
         },
         rewriter);
   }
 };
 
 // A `rsqrt` is converted into `1 / sqrt`.
 struct RsqrtOpLowering : public ConvertOpToLLVMPattern<math::RsqrtOp> {
   using ConvertOpToLLVMPattern<math::RsqrtOp>::ConvertOpToLLVMPattern;
 
   LogicalResult
   matchAndRewrite(math::RsqrtOp op, OpAdaptor adaptor,
                   ConversionPatternRewriter &rewriter) const override {
     auto operandType = adaptor.getOperand().getType();
 
     if (!operandType || !LLVM::isCompatibleType(operandType))
       return failure();
 
     auto loc = op.getLoc();
     auto resultType = op.getResult().getType();
     auto floatType = getElementTypeOrSelf(resultType).cast<FloatType>();
     auto floatOne = rewriter.getFloatAttr(floatType, 1.0);
+    LLVM::detail::ConvertToLLVMAttributes sqrtAttrs(
+        op, LLVM::SqrtOp::getOperationName());
+    LLVM::detail::ConvertToLLVMAttributes divAttrs(
+        op, LLVM::FDivOp::getOperationName());
 
     if (!operandType.isa<LLVM::LLVMArrayType>()) {
       LLVM::ConstantOp one;
       if (LLVM::isCompatibleVectorType(operandType)) {
         one = rewriter.create<LLVM::ConstantOp>(
             loc, operandType,
             SplatElementsAttr::get(resultType.cast<ShapedType>(), floatOne));
       } else {
         one = rewriter.create<LLVM::ConstantOp>(loc, operandType, floatOne);
       }
-      auto sqrt = rewriter.create<LLVM::SqrtOp>(loc, adaptor.getOperand());
-      rewriter.replaceOpWithNewOp<LLVM::FDivOp>(op, operandType, one, sqrt);
+      auto sqrt = rewriter.create<LLVM::SqrtOp>(loc, adaptor.getOperand(),
+                                                sqrtAttrs.getAttrs());
+      rewriter.replaceOpWithNewOp<LLVM::FDivOp>(
+          op, operandType, ValueRange{one, sqrt}, divAttrs.getAttrs());
       return success();
     }
 
     auto vectorType = resultType.dyn_cast<VectorType>();
     if (!vectorType)
       return failure();
 
     return LLVM::detail::handleMultidimensionalVectors(
         op.getOperation(), adaptor.getOperands(), *getTypeConverter(),
         [&](Type llvm1DVectorTy, ValueRange operands) {
           auto splatAttr = SplatElementsAttr::get(
               mlir::VectorType::get(
                   {LLVM::getVectorNumElements(llvm1DVectorTy).getFixedValue()},
                   floatType),
               floatOne);
           auto one =
               rewriter.create<LLVM::ConstantOp>(loc, llvm1DVectorTy, splatAttr);
-          auto sqrt =
-              rewriter.create<LLVM::SqrtOp>(loc, llvm1DVectorTy, operands[0]);
-          return rewriter.create<LLVM::FDivOp>(loc, llvm1DVectorTy, one, sqrt);
+          auto sqrt = rewriter.create<LLVM::SqrtOp>(
+              loc, llvm1DVectorTy, operands[0], sqrtAttrs.getAttrs());
+          return rewriter.create<LLVM::FDivOp>(
+              loc, llvm1DVectorTy, ValueRange{one, sqrt}, divAttrs.getAttrs());
         },
         rewriter);
   }
 };
 
 struct ConvertMathToLLVMPass
     : public impl::ConvertMathToLLVMBase<ConvertMathToLLVMPass> {
   ConvertMathToLLVMPass() = default;

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

 add_mlir_dialect_library(MLIRMathDialect
   MathOps.cpp
   MathDialect.cpp
 
   ADDITIONAL_HEADER_DIRS
   ${MLIR_MAIN_INCLUDE_DIR}/mlir/Dialect/Math
 
   DEPENDS
+  MLIRMathOpsAttributesIncGen
   MLIRMathOpsIncGen
+  MLIRMathOpsInterfacesIncGen
 
   LINK_LIBS PUBLIC
   MLIRArithDialect
   MLIRDialect
   MLIRIR
   )

mlir/lib/Dialect/Math/IR/MathDialect.cpp

 //===- MathDialect.cpp - MLIR dialect for Math implementation -------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #include "mlir/Dialect/Math/IR/Math.h"
 #include "mlir/Transforms/InliningUtils.h"
 
+#include "mlir/IR/DialectImplementation.h"
+#include "llvm/ADT/TypeSwitch.h"
+
 using namespace mlir;
 using namespace mlir::math;
 
 #include "mlir/Dialect/Math/IR/MathOpsDialect.cpp.inc"
+#include "mlir/Dialect/Math/IR/MathOpsInterfaces.cpp.inc"
+#define GET_ATTRDEF_CLASSES
+#include "mlir/Dialect/Math/IR/MathOpsAttributes.cpp.inc"
 
 namespace {
 /// This class defines the interface for handling inlining with math
 /// operations.
 struct MathInlinerInterface : public DialectInlinerInterface {
   using DialectInlinerInterface::DialectInlinerInterface;
 
   /// All operations within math ops can be inlined.
   bool isLegalToInline(Operation *, Region *, bool,
                        BlockAndValueMapping &) const final {
     return true;
   }
 };
 } // namespace
 
 void mlir::math::MathDialect::initialize() {
   addOperations<
 #define GET_OP_LIST
 #include "mlir/Dialect/Math/IR/MathOps.cpp.inc"
       >();
+  addAttributes<
+#define GET_ATTRDEF_LIST
+#include "mlir/Dialect/Math/IR/MathOpsAttributes.cpp.inc"
+      >();
   addInterfaces<MathInlinerInterface>();
 }

mlir/lib/Dialect/Math/IR/MathOps.cpp

 #include "mlir/Dialect/CommonFolders.h"
 #include "mlir/Dialect/Math/IR/Math.h"
 #include "mlir/IR/Builders.h"
 
 using namespace mlir;
 using namespace mlir::math;
 
 //===----------------------------------------------------------------------===//
+// Floating point op parse/print helpers
+//===----------------------------------------------------------------------===//
+static ParseResult parseMathFastMathAttr(OpAsmParser &parser, Attribute &attr) {
+  if (succeeded(
+          parser.parseOptionalKeyword(FastMathFlagsAttr::getMnemonic()))) {
+    attr = FastMathFlagsAttr::parse(parser, Type{});
+    return success(static_cast<bool>(attr));
+  } else {
+    // No fastmath attribute mnemonic present - defer attribute creation and use
+    // the default value.
+    return success();
+  }
+}
+
+static void printMathFastMathAttr(OpAsmPrinter &printer, Operation *op,
+                                  FastMathFlagsAttr fmAttr) {
+  // Elide printing the fastmath attribute when fastmath=none
+  if (fmAttr && (fmAttr.getValue() != FastMathFlags::none)) {
+    printer << " " << FastMathFlagsAttr::getMnemonic();
+    fmAttr.print(printer);
+  }
+}
+
+//===----------------------------------------------------------------------===//
 // TableGen'd op method definitions
 //===----------------------------------------------------------------------===//
 
 #define GET_OP_CLASSES
 #include "mlir/Dialect/Math/IR/MathOps.cpp.inc"
 
 //===----------------------------------------------------------------------===//
+// TableGen'd enum attribute definitions
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Dialect/Math/IR/MathOpsEnums.cpp.inc"
+
+//===----------------------------------------------------------------------===//
 // AbsFOp folder
 //===----------------------------------------------------------------------===//
 
 OpFoldResult math::AbsFOp::fold(ArrayRef<Attribute> operands) {
   return constFoldUnaryOp<FloatAttr>(operands,
                                      [](const APFloat &a) { return abs(a); });
 }
 

mlir/test/Conversion/MathToLLVM/math-to-llvm.mlir

func.func @log1p(%arg0 : f32) {
   // CHECK: %[[ADD:.*]] = llvm.fadd %[[ONE]], %arg0 : f32
   // CHECK: %[[LOG:.*]] = llvm.intr.log(%[[ADD]]) : (f32) -> f32
   %0 = math.log1p %arg0 : f32
   func.return
 }
 
 // -----
 
+// CHECK-LABEL: func @log1p_fmf(
+// CHECK-SAME: f32
+func.func @log1p_fmf(%arg0 : f32) {
+  // CHECK: %[[ONE:.*]] = llvm.mlir.constant(1.000000e+00 : f32) : f32
+  // CHECK: %[[ADD:.*]] = llvm.fadd %[[ONE]], %arg0 {fastmathFlags = #llvm.fastmath<fast>} : f32
+  // CHECK: %[[LOG:.*]] = llvm.intr.log(%[[ADD]]) {fastmathFlags = #llvm.fastmath<fast>} : (f32) -> f32
+  %0 = math.log1p %arg0 fastmath<fast> : f32
+  func.return
+}
+
+// -----
+
 // CHECK-LABEL: func @log1p_2dvector(
 func.func @log1p_2dvector(%arg0 : vector<4x3xf32>) {
   // CHECK: %[[EXTRACT:.*]] = llvm.extractvalue %{{.*}}[0] : !llvm.array<4 x vector<3xf32>>
   // CHECK: %[[ONE:.*]] = llvm.mlir.constant(dense<1.000000e+00> : vector<3xf32>) : vector<3xf32>
   // CHECK: %[[ADD:.*]] = llvm.fadd %[[ONE]], %[[EXTRACT]] : vector<3xf32>
   // CHECK: %[[LOG:.*]] = llvm.intr.log(%[[ADD]]) : (vector<3xf32>) -> vector<3xf32>
   // CHECK: %[[INSERT:.*]] = llvm.insertvalue %[[LOG]], %{{.*}}[0] : !llvm.array<4 x vector<3xf32>>
   %0 = math.log1p %arg0 : vector<4x3xf32>
   func.return
 }
 
 // -----
 
+// CHECK-LABEL: func @log1p_2dvector_fmf(
+func.func @log1p_2dvector_fmf(%arg0 : vector<4x3xf32>) {
+  // CHECK: %[[EXTRACT:.*]] = llvm.extractvalue %{{.*}}[0] : !llvm.array<4 x vector<3xf32>>
+  // CHECK: %[[ONE:.*]] = llvm.mlir.constant(dense<1.000000e+00> : vector<3xf32>) : vector<3xf32>
+  // CHECK: %[[ADD:.*]] = llvm.fadd %[[ONE]], %[[EXTRACT]] {fastmathFlags = #llvm.fastmath<fast>} : vector<3xf32>
+  // CHECK: %[[LOG:.*]] = llvm.intr.log(%[[ADD]]) {fastmathFlags = #llvm.fastmath<fast>} : (vector<3xf32>) -> vector<3xf32>
+  // CHECK: %[[INSERT:.*]] = llvm.insertvalue %[[LOG]], %{{.*}}[0] : !llvm.array<4 x vector<3xf32>>
+  %0 = math.log1p %arg0 fastmath<fast> : vector<4x3xf32>
+  func.return
+}
+
+// -----
+
 // CHECK-LABEL: func @expm1(
 // CHECK-SAME: f32
 func.func @expm1(%arg0 : f32) {
   // CHECK: %[[ONE:.*]] = llvm.mlir.constant(1.000000e+00 : f32) : f32
   // CHECK: %[[EXP:.*]] = llvm.intr.exp(%arg0) : (f32) -> f32
   // CHECK: %[[SUB:.*]] = llvm.fsub %[[EXP]], %[[ONE]] : f32
   %0 = math.expm1 %arg0 : f32
   func.return
 }
 
 // -----
 
+// CHECK-LABEL: func @expm1_fmf(
+// CHECK-SAME: f32
+func.func @expm1_fmf(%arg0 : f32) {
+  // CHECK: %[[ONE:.*]] = llvm.mlir.constant(1.000000e+00 : f32) : f32
+  // CHECK: %[[EXP:.*]] = llvm.intr.exp(%arg0) {fastmathFlags = #llvm.fastmath<fast>} : (f32) -> f32
+  // CHECK: %[[SUB:.*]] = llvm.fsub %[[EXP]], %[[ONE]] {fastmathFlags = #llvm.fastmath<fast>} : f32
+  %0 = math.expm1 %arg0 fastmath<fast> : f32
+  func.return
+}
+
+// -----
+
+// CHECK-LABEL: func @expm1_vector(
+// CHECK-SAME: vector<4xf32>
+func.func @expm1_vector(%arg0 : vector<4xf32>) {
+  // CHECK: %[[ONE:.*]] = llvm.mlir.constant(dense<1.000000e+00> : vector<4xf32>) : vector<4xf32>
+  // CHECK: %[[EXP:.*]] = llvm.intr.exp(%arg0) : (vector<4xf32>) -> vector<4xf32>
+  // CHECK: %[[SUB:.*]] = llvm.fsub %[[EXP]], %[[ONE]] : vector<4xf32>
+  %0 = math.expm1 %arg0 : vector<4xf32>
+  func.return
+}
+
+// -----
+
+// CHECK-LABEL: func @expm1_vector_fmf(
+// CHECK-SAME: vector<4xf32>
+func.func @expm1_vector_fmf(%arg0 : vector<4xf32>) {
+  // CHECK: %[[ONE:.*]] = llvm.mlir.constant(dense<1.000000e+00> : vector<4xf32>) : vector<4xf32>
+  // CHECK: %[[EXP:.*]] = llvm.intr.exp(%arg0) {fastmathFlags = #llvm.fastmath<fast>} : (vector<4xf32>) -> vector<4xf32>
+  // CHECK: %[[SUB:.*]] = llvm.fsub %[[EXP]], %[[ONE]] {fastmathFlags = #llvm.fastmath<fast>} : vector<4xf32>
+  %0 = math.expm1 %arg0 fastmath<fast> : vector<4xf32>
+  func.return
+}
+
+// -----
+
 // CHECK-LABEL: func @rsqrt(
 // CHECK-SAME: f32
 func.func @rsqrt(%arg0 : f32) {
   // CHECK: %[[ONE:.*]] = llvm.mlir.constant(1.000000e+00 : f32) : f32
   // CHECK: %[[SQRT:.*]] = llvm.intr.sqrt(%arg0) : (f32) -> f32
   // CHECK: %[[DIV:.*]] = llvm.fdiv %[[ONE]], %[[SQRT]] : f32
   %0 = math.rsqrt %arg0 : f32
   func.return
func.func @rsqrt_double(%arg0 : f64) {
   // CHECK: %[[SQRT:.*]] = llvm.intr.sqrt(%arg0) : (f64) -> f64
   // CHECK: %[[DIV:.*]] = llvm.fdiv %[[ONE]], %[[SQRT]] : f64
   %0 = math.rsqrt %arg0 : f64
   func.return
 }
 
 // -----
 
+// CHECK-LABEL: func @rsqrt_double_fmf(
+// CHECK-SAME: f64
+func.func @rsqrt_double_fmf(%arg0 : f64) {
+  // CHECK: %[[ONE:.*]] = llvm.mlir.constant(1.000000e+00 : f64) : f64
+  // CHECK: %[[SQRT:.*]] = llvm.intr.sqrt(%arg0) {fastmathFlags = #llvm.fastmath<fast>} : (f64) -> f64
+  // CHECK: %[[DIV:.*]] = llvm.fdiv %[[ONE]], %[[SQRT]] {fastmathFlags = #llvm.fastmath<fast>} : f64
+  %0 = math.rsqrt %arg0 fastmath<fast> : f64
+  func.return
+}
+
+// -----
+
 // CHECK-LABEL: func @rsqrt_vector(
 // CHECK-SAME: vector<4xf32>
 func.func @rsqrt_vector(%arg0 : vector<4xf32>) {
   // CHECK: %[[ONE:.*]] = llvm.mlir.constant(dense<1.000000e+00> : vector<4xf32>) : vector<4xf32>
   // CHECK: %[[SQRT:.*]] = llvm.intr.sqrt(%arg0) : (vector<4xf32>) -> vector<4xf32>
   // CHECK: %[[DIV:.*]] = llvm.fdiv %[[ONE]], %[[SQRT]] : vector<4xf32>
   %0 = math.rsqrt %arg0 : vector<4xf32>
   func.return
 }
 
 // -----
 
+// CHECK-LABEL: func @rsqrt_vector_fmf(
+// CHECK-SAME: vector<4xf32>
+func.func @rsqrt_vector_fmf(%arg0 : vector<4xf32>) {
+  // CHECK: %[[ONE:.*]] = llvm.mlir.constant(dense<1.000000e+00> : vector<4xf32>) : vector<4xf32>
+  // CHECK: %[[SQRT:.*]] = llvm.intr.sqrt(%arg0) {fastmathFlags = #llvm.fastmath<fast>} : (vector<4xf32>) -> vector<4xf32>
+  // CHECK: %[[DIV:.*]] = llvm.fdiv %[[ONE]], %[[SQRT]] {fastmathFlags = #llvm.fastmath<fast>} : vector<4xf32>
+  %0 = math.rsqrt %arg0 fastmath<fast> : vector<4xf32>
+  func.return
+}
+
+// -----
+
 // CHECK-LABEL: func @rsqrt_multidim_vector(
 func.func @rsqrt_multidim_vector(%arg0 : vector<4x3xf32>) {
   // CHECK: %[[EXTRACT:.*]] = llvm.extractvalue %{{.*}}[0] : !llvm.array<4 x vector<3xf32>>
   // CHECK: %[[ONE:.*]] = llvm.mlir.constant(dense<1.000000e+00> : vector<3xf32>) : vector<3xf32>
   // CHECK: %[[SQRT:.*]] = llvm.intr.sqrt(%[[EXTRACT]]) : (vector<3xf32>) -> vector<3xf32>
   // CHECK: %[[DIV:.*]] = llvm.fdiv %[[ONE]], %[[SQRT]] : vector<3xf32>
   // CHECK: %[[INSERT:.*]] = llvm.insertvalue %[[DIV]], %{{.*}}[0] : !llvm.array<4 x vector<3xf32>>
   %0 = math.rsqrt %arg0 : vector<4x3xf32>

 
 // CHECK-LABEL: func @trunc(
 // CHECK-SAME: f32
 func.func @trunc(%arg0 : f32) {
   // CHECK: llvm.intr.trunc(%arg0) : (f32) -> f32
   %0 = math.trunc %arg0 : f32
   func.return
 }
+
+// -----
+
+// CHECK-LABEL: func @fastmath(
+// CHECK-SAME: f32
+func.func @fastmath(%arg0 : f32, %arg1 : vector<4xf32>) {
+  // CHECK: llvm.intr.trunc(%arg0) {fastmathFlags = #llvm.fastmath<fast>} : (f32) -> f32
+  %0 = math.trunc %arg0 fastmath<fast> : f32
+  // CHECK: llvm.intr.pow(%arg0, %arg0) {fastmathFlags = #llvm.fastmath<afn>} : (f32, f32) -> f32
+  %1 = math.powf %arg0, %arg0 fastmath<afn> : f32
+  // CHECK: llvm.intr.sqrt(%arg0) : (f32) -> f32
+  %2 = math.sqrt %arg0 fastmath<none> : f32
+  // CHECK: llvm.intr.fma(%arg0, %arg0, %arg0) {fastmathFlags = #llvm.fastmath<fast>} : (f32, f32, f32) -> f32
+  %3 = math.fma %arg0, %arg0, %arg0 fastmath<reassoc,nnan,ninf,nsz,arcp,contract,afn> : f32
+  func.return
+}

mlir/test/Dialect/Math/ops.mlir

func.func @trunc(%f: f32, %v: vector<4xf32>, %t: tensor<4x4x?xf32>) {
   // CHECK: %{{.*}} = math.trunc %[[F]] : f32
   %0 = math.trunc %f : f32
   // CHECK: %{{.*}} = math.trunc %[[V]] : vector<4xf32>
   %1 = math.trunc %v : vector<4xf32>
   // CHECK: %{{.*}} = math.trunc %[[T]] : tensor<4x4x?xf32>
   %2 = math.trunc %t : tensor<4x4x?xf32>
   return
 }
+
+// CHECK-LABEL: func @fastmath(
+// CHECK-SAME:             %[[F:.*]]: f32, %[[V:.*]]: vector<4xf32>, %[[T:.*]]: tensor<4x4x?xf32>)
+func.func @fastmath(%f: f32, %v: vector<4xf32>, %t: tensor<4x4x?xf32>) {
+  // CHECK: %{{.*}} = math.trunc %[[F]] fastmath<fast> : f32
+  %0 = math.trunc %f fastmath<fast> : f32
+  // CHECK: %{{.*}} = math.powf %[[V]], %[[V]] fastmath<fast> : vector<4xf32>
+  %1 = math.powf %v, %v fastmath<reassoc,nnan,ninf,nsz,arcp,contract,afn> : vector<4xf32>
+  // CHECK: %{{.*}} = math.fma %[[T]], %[[T]], %[[T]] : tensor<4x4x?xf32>
+  %2 = math.fma %t, %t, %t fastmath<none> : tensor<4x4x?xf32>
+  return
+}