[MLIR][LLVM] Add llvm.experimental.constrained.fptrunc operation (#86260)

Add operation mapping to the LLVM
`llvm.experimental.constrained.fptrunc.*` intrinsic.

The new operation implements the new
`LLVM::FPExceptionBehaviorOpInterface` and
`LLVM::RoundingModeOpInterface` interfaces.

---------

Signed-off-by: Victor Perez <victor.perez@codeplay.com>

mlir/include/mlir/Dialect/LLVMIR/LLVMEnums.td

@@ -705,4 +705,61 @@ def FramePointerKindEnum : LLVM_EnumAttr<
   let cppNamespace = "::mlir::LLVM::framePointerKind";
 }
 
+//===----------------------------------------------------------------------===//
+// RoundingMode
+//===----------------------------------------------------------------------===//
+
+// These values must match llvm::RoundingMode ones.
+// See llvm/include/llvm/ADT/FloatingPointMode.h.
+def RoundTowardZero
+    : LLVM_EnumAttrCase<"TowardZero", "towardzero", "TowardZero", 0>;
+def RoundNearestTiesToEven
+    : LLVM_EnumAttrCase<"NearestTiesToEven", "tonearest", "NearestTiesToEven", 1>;
+def RoundTowardPositive
+    : LLVM_EnumAttrCase<"TowardPositive", "upward", "TowardPositive", 2>;
+def RoundTowardNegative
+    : LLVM_EnumAttrCase<"TowardNegative", "downward", "TowardNegative", 3>;
+def RoundNearestTiesToAway
+    : LLVM_EnumAttrCase<"NearestTiesToAway", "tonearestaway", "NearestTiesToAway", 4>;
+def RoundDynamic
+    : LLVM_EnumAttrCase<"Dynamic", "dynamic", "Dynamic", 7>;
+// Needed as llvm::RoundingMode defines this.
+def RoundInvalid
+    : LLVM_EnumAttrCase<"Invalid", "invalid", "Invalid", -1>;
+
+// RoundingModeAttr should not be used in operations definitions.
+// Use ValidRoundingModeAttr instead.
+def RoundingModeAttr : LLVM_EnumAttr<
+    "RoundingMode",
+    "::llvm::RoundingMode",
+    "LLVM Rounding Mode",
+    [RoundTowardZero, RoundNearestTiesToEven, RoundTowardPositive,
+     RoundTowardNegative, RoundNearestTiesToAway, RoundDynamic, RoundInvalid]> {
+  let cppNamespace = "::mlir::LLVM";
+}
+
+def ValidRoundingModeAttr : ConfinedAttr<RoundingModeAttr, [IntMinValue<0>]>;
+
+//===----------------------------------------------------------------------===//
+// FPExceptionBehavior
+//===----------------------------------------------------------------------===//
+
+// These values must match llvm::fp::ExceptionBehavior ones.
+// See llvm/include/llvm/IR/FPEnv.h.
+def FPExceptionBehaviorIgnore
+    : LLVM_EnumAttrCase<"Ignore", "ignore", "ebIgnore", 0>;
+def FPExceptionBehaviorMayTrap
+    : LLVM_EnumAttrCase<"MayTrap", "maytrap", "ebMayTrap", 1>;
+def FPExceptionBehaviorStrict
+    : LLVM_EnumAttrCase<"Strict", "strict", "ebStrict", 2>;
+
+def FPExceptionBehaviorAttr : LLVM_EnumAttr<
+    "FPExceptionBehavior",
+    "::llvm::fp::ExceptionBehavior",
+    "LLVM Exception Behavior",
+    [FPExceptionBehaviorIgnore, FPExceptionBehaviorMayTrap,
+     FPExceptionBehaviorStrict]> {
+  let cppNamespace = "::mlir::LLVM";
+}
+
 #endif // LLVMIR_ENUMS

mlir/include/mlir/Dialect/LLVMIR/LLVMInterfaces.td

@@ -290,6 +290,73 @@ def GetResultPtrElementType : OpInterface<"GetResultPtrElementType"> {
   ];
 }
 
+def FPExceptionBehaviorOpInterface : OpInterface<"FPExceptionBehaviorOpInterface"> {
+  let description = [{
+    An interface for operations receiving an exception behavior attribute
+    controlling FP exception behavior.
+  }];
+
+  let cppNamespace = "::mlir::LLVM";
+
+  let methods = [
+    InterfaceMethod<
+      /*desc=*/        "Returns a FPExceptionBehavior attribute for the operation",
+      /*returnType=*/  "FPExceptionBehaviorAttr",
+      /*methodName=*/  "getFPExceptionBehaviorAttr",
+      /*args=*/        (ins),
+      /*methodBody=*/  [{}],
+      /*defaultImpl=*/ [{
+        auto op = cast<ConcreteOp>(this->getOperation());
+        return op.getFpExceptionBehaviorAttr();
+      }]
+    >,
+    StaticInterfaceMethod<
+      /*desc=*/        [{Returns the name of the FPExceptionBehaviorAttr
+                        attribute for the operation}],
+      /*returnType=*/  "StringRef",
+      /*methodName=*/  "getFPExceptionBehaviorAttrName",
+      /*args=*/        (ins),
+      /*methodBody=*/  [{}],
+      /*defaultImpl=*/ [{
+        return "fpExceptionBehavior";
+      }]
+    >
+  ];
+}
+
+def RoundingModeOpInterface : OpInterface<"RoundingModeOpInterface"> {
+  let description = [{
+    An interface for operations receiving a rounding mode attribute
+    controlling FP rounding mode.
+  }];
+
+  let cppNamespace = "::mlir::LLVM";
+
+  let methods = [
+    InterfaceMethod<
+      /*desc=*/        "Returns a RoundingMode attribute for the operation",
+      /*returnType=*/  "RoundingModeAttr",
+      /*methodName=*/  "getRoundingModeAttr",
+      /*args=*/        (ins),
+      /*methodBody=*/  [{}],
+      /*defaultImpl=*/ [{
+        auto op = cast<ConcreteOp>(this->getOperation());
+        return op.getRoundingmodeAttr();
+      }]
+    >,
+    StaticInterfaceMethod<
+      /*desc=*/        [{Returns the name of the RoundingModeAttr attribute
+                         for the operation}],
+      /*returnType=*/  "StringRef",
+      /*methodName=*/  "getRoundingModeAttrName",
+      /*args=*/        (ins),
+      /*methodBody=*/  [{}],
+      /*defaultImpl=*/ [{
+        return "roundingmode";
+      }]
+    >,
+  ];
+}
 
 //===----------------------------------------------------------------------===//
 // LLVM dialect type interfaces.

mlir/include/mlir/Dialect/LLVMIR/LLVMIntrinsicOps.td

@@ -311,6 +311,91 @@ def LLVM_InvariantEndOp : LLVM_ZeroResultIntrOp<"invariant.end", [2],
       "qualified(type($ptr))";
 }
 
+// Constrained Floating-Point Intrinsics.
+
+class LLVM_ConstrainedIntr<string mnem, int numArgs,
+                           bit overloadedResult, list<int> overloadedOperands,
+                           bit hasRoundingMode>
+    : LLVM_OneResultIntrOp<"experimental.constrained." # mnem,
+                           /*overloadedResults=*/
+                           !cond(!gt(overloadedResult, 0) : [0],
+                                 true : []),
+                           overloadedOperands,
+                           /*traits=*/[Pure, DeclareOpInterfaceMethods<FPExceptionBehaviorOpInterface>]
+                           # !cond(
+                               !gt(hasRoundingMode, 0) : [DeclareOpInterfaceMethods<RoundingModeOpInterface>],
+                               true : []),
+                           /*requiresFastmath=*/0,
+                           /*immArgPositions=*/[],
+                           /*immArgAttrNames=*/[]> {
+  dag regularArgs = !dag(ins, !listsplat(LLVM_Type, numArgs), !foreach(i, !range(numArgs), "arg_" #i));
+  dag attrArgs = !con(!cond(!gt(hasRoundingMode, 0) : (ins ValidRoundingModeAttr:$roundingmode),
+                            true : (ins)),
+                      (ins FPExceptionBehaviorAttr:$fpExceptionBehavior));
+  let arguments = !con(regularArgs, attrArgs);
+  let llvmBuilder = [{
+    SmallVector<llvm::Value *> args =
+      moduleTranslation.lookupValues(opInst.getOperands());
+    SmallVector<llvm::Type *> overloadedTypes; }] #
+    !cond(!gt(overloadedResult, 0) : [{
+    // Take into account overloaded result type.
+    overloadedTypes.push_back($_resultType); }],
+    // No overloaded result type.
+          true : "") # [{
+    llvm::transform(ArrayRef<unsigned>}] # overloadedOperandsCpp # [{,
+                    std::back_inserter(overloadedTypes),
+                    [&args](unsigned index) { return args[index]->getType(); });
+    llvm::Module *module = builder.GetInsertBlock()->getModule();
+    llvm::Function *callee =
+      llvm::Intrinsic::getDeclaration(module,
+        llvm::Intrinsic::experimental_constrained_}] #
+    mnem # [{, overloadedTypes); }] #
+    !cond(!gt(hasRoundingMode, 0) : [{
+    // Get rounding mode using interface.
+    llvm::RoundingMode rounding =
+        moduleTranslation.translateRoundingMode($roundingmode); }],
+          true : [{
+    // No rounding mode.
+    std::optional<llvm::RoundingMode> rounding; }]) # [{
+    llvm::fp::ExceptionBehavior except =
+      moduleTranslation.translateFPExceptionBehavior($fpExceptionBehavior);
+    $res = builder.CreateConstrainedFPCall(callee, args, "", rounding, except);
+  }];
+  let mlirBuilder = [{
+    SmallVector<Value> mlirOperands;
+    SmallVector<NamedAttribute> mlirAttrs;
+    if (failed(moduleImport.convertIntrinsicArguments(
+        llvmOperands.take_front( }] # numArgs # [{),
+        {}, {}, mlirOperands, mlirAttrs))) {
+      return failure();
+    }
+
+    FPExceptionBehaviorAttr fpExceptionBehaviorAttr =
+        $_fpExceptionBehavior_attr($fpExceptionBehavior);
+    mlirAttrs.push_back(
+        $_builder.getNamedAttr(
+            $_qualCppClassName::getFPExceptionBehaviorAttrName(),
+            fpExceptionBehaviorAttr)); }] #
+    !cond(!gt(hasRoundingMode, 0) : [{
+    RoundingModeAttr roundingModeAttr = $_roundingMode_attr($roundingmode);
+    mlirAttrs.push_back(
+        $_builder.getNamedAttr($_qualCppClassName::getRoundingModeAttrName(),
+                               roundingModeAttr));
+    }], true : "") # [{
+    $res = $_builder.create<$_qualCppClassName>($_location,
+      $_resultType, mlirOperands, mlirAttrs);
+  }];
+}
+
+def LLVM_ConstrainedFPTruncIntr
+    : LLVM_ConstrainedIntr<"fptrunc", /*numArgs=*/1,
+        /*overloadedResult=*/1, /*overloadedOperands=*/[0],
+        /*hasRoundingMode=*/1> {
+  let assemblyFormat = [{
+    $arg_0 $roundingmode $fpExceptionBehavior attr-dict `:` type($arg_0) `to` type(results)
+  }];
+}
+
 // Intrinsics with multiple returns.
 
 class LLVM_ArithWithOverflowOp<string mnem>

mlir/include/mlir/Dialect/LLVMIR/LLVMOpBase.td

@@ -170,6 +170,10 @@ class LLVM_OpBase<Dialect dialect, string mnemonic, list<Trait> traits = []> :
   //   - $_float_attr - substituted by a call to a float attribute matcher;
   //   - $_var_attr - substituted by a call to a variable attribute matcher;
   //   - $_label_attr - substituted by a call to a label attribute matcher;
+  //   - $_roundingMode_attr - substituted by a call to a rounding mode
+  //     attribute matcher;
+  //   - $_fpExceptionBehavior_attr - substituted by a call to a FP exception
+  //     behavior attribute matcher;
   //   - $_resultType - substituted with the MLIR result type;
   //   - $_location - substituted with the MLIR location;
   //   - $_builder - substituted with the MLIR builder;

mlir/include/mlir/Target/LLVMIR/ModuleImport.h

@@ -152,6 +152,14 @@ public:
   /// Converts `value` to a label attribute. Asserts if the matching fails.
   DILabelAttr matchLabelAttr(llvm::Value *value);
 
+  /// Converts `value` to a FP exception behavior attribute. Asserts if the
+  /// matching fails.
+  FPExceptionBehaviorAttr matchFPExceptionBehaviorAttr(llvm::Value *value);
+
+  /// Converts `value` to a rounding mode attribute. Asserts if the matching
+  /// fails.
+  RoundingModeAttr matchRoundingModeAttr(llvm::Value *value);
+
   /// Converts `value` to an array of alias scopes or returns failure if the
   /// conversion fails.
   FailureOr<SmallVector<AliasScopeAttr>>

mlir/include/mlir/Target/LLVMIR/ModuleTranslation.h

@@ -201,6 +201,13 @@ public:
   /// Translates the given LLVM debug info metadata.
   llvm::Metadata *translateDebugInfo(LLVM::DINodeAttr attr);
 
+  /// Translates the given LLVM rounding mode metadata.
+  llvm::RoundingMode translateRoundingMode(LLVM::RoundingMode rounding);
+
+  /// Translates the given LLVM FP exception behavior metadata.
+  llvm::fp::ExceptionBehavior
+  translateFPExceptionBehavior(LLVM::FPExceptionBehavior exceptionBehavior);
+
   /// Translates the contents of the given block to LLVM IR using this
   /// translator. The LLVM IR basic block corresponding to the given block is
   /// expected to exist in the mapping of this translator. Uses `builder` to

mlir/lib/Target/LLVMIR/ModuleImport.cpp

@@ -1290,6 +1290,27 @@ DILabelAttr ModuleImport::matchLabelAttr(llvm::Value *value) {
   return debugImporter->translate(node);
 }
 
+FPExceptionBehaviorAttr
+ModuleImport::matchFPExceptionBehaviorAttr(llvm::Value *value) {
+  auto *metadata = cast<llvm::MetadataAsValue>(value);
+  auto *mdstr = cast<llvm::MDString>(metadata->getMetadata());
+  std::optional<llvm::fp::ExceptionBehavior> optLLVM =
+      llvm::convertStrToExceptionBehavior(mdstr->getString());
+  assert(optLLVM && "Expecting FP exception behavior");
+  return builder.getAttr<FPExceptionBehaviorAttr>(
+      convertFPExceptionBehaviorFromLLVM(*optLLVM));
+}
+
+RoundingModeAttr ModuleImport::matchRoundingModeAttr(llvm::Value *value) {
+  auto *metadata = cast<llvm::MetadataAsValue>(value);
+  auto *mdstr = cast<llvm::MDString>(metadata->getMetadata());
+  std::optional<llvm::RoundingMode> optLLVM =
+      llvm::convertStrToRoundingMode(mdstr->getString());
+  assert(optLLVM && "Expecting rounding mode");
+  return builder.getAttr<RoundingModeAttr>(
+      convertRoundingModeFromLLVM(*optLLVM));
+}
+
 FailureOr<SmallVector<AliasScopeAttr>>
 ModuleImport::matchAliasScopeAttrs(llvm::Value *value) {
   auto *nodeAsVal = cast<llvm::MetadataAsValue>(value);

mlir/lib/Target/LLVMIR/ModuleTranslation.cpp

@@ -1721,6 +1721,16 @@ llvm::Metadata *ModuleTranslation::translateDebugInfo(LLVM::DINodeAttr attr) {
   return debugTranslation->translate(attr);
 }
 
+llvm::RoundingMode
+ModuleTranslation::translateRoundingMode(LLVM::RoundingMode rounding) {
+  return convertRoundingModeToLLVM(rounding);
+}
+
+llvm::fp::ExceptionBehavior ModuleTranslation::translateFPExceptionBehavior(
+    LLVM::FPExceptionBehavior exceptionBehavior) {
+  return convertFPExceptionBehaviorToLLVM(exceptionBehavior);
+}
+
 llvm::NamedMDNode *
 ModuleTranslation::getOrInsertNamedModuleMetadata(StringRef name) {
   return llvmModule->getOrInsertNamedMetadata(name);

mlir/test/Dialect/LLVMIR/roundtrip.mlir

@@ -647,3 +647,18 @@ llvm.func @experimental_noalias_scope_decl() {
   llvm.intr.experimental.noalias.scope.decl #alias_scope
   llvm.return
 }
+
+// CHECK-LABEL: @experimental_constrained_fptrunc
+llvm.func @experimental_constrained_fptrunc(%in: f64) {
+  // CHECK: llvm.intr.experimental.constrained.fptrunc %{{.*}} towardzero ignore : f64 to f32
+  %0 = llvm.intr.experimental.constrained.fptrunc %in towardzero ignore : f64 to f32
+  // CHECK: llvm.intr.experimental.constrained.fptrunc %{{.*}} tonearest maytrap : f64 to f32
+  %1 = llvm.intr.experimental.constrained.fptrunc %in tonearest maytrap : f64 to f32
+  // CHECK: llvm.intr.experimental.constrained.fptrunc %{{.*}} upward strict : f64 to f32
+  %2 = llvm.intr.experimental.constrained.fptrunc %in upward strict : f64 to f32
+  // CHECK: llvm.intr.experimental.constrained.fptrunc %{{.*}} downward ignore : f64 to f32
+  %3 = llvm.intr.experimental.constrained.fptrunc %in downward ignore : f64 to f32
+  // CHECK: llvm.intr.experimental.constrained.fptrunc %{{.*}} tonearestaway ignore : f64 to f32
+  %4 = llvm.intr.experimental.constrained.fptrunc %in tonearestaway ignore : f64 to f32
+  llvm.return
+}

mlir/test/Target/LLVMIR/Import/intrinsic.ll

@@ -894,6 +894,23 @@ define float @ssa_copy(float %0) {
   ret float %2
 }
 
+; CHECK-LABEL: experimental_constrained_fptrunc
+define void @experimental_constrained_fptrunc(double %s, <4 x double> %v) {
+  ; CHECK: llvm.intr.experimental.constrained.fptrunc %{{.*}} towardzero ignore : f64 to f32
+  %1 = call float @llvm.experimental.constrained.fptrunc.f32.f64(double %s, metadata !"round.towardzero", metadata !"fpexcept.ignore")
+  ; CHECK: llvm.intr.experimental.constrained.fptrunc %{{.*}} tonearest maytrap : f64 to f32
+  %2 = call float @llvm.experimental.constrained.fptrunc.f32.f64(double %s, metadata !"round.tonearest", metadata !"fpexcept.maytrap")
+  ; CHECK: llvm.intr.experimental.constrained.fptrunc %{{.*}} upward strict : f64 to f32
+  %3 = call float @llvm.experimental.constrained.fptrunc.f32.f64(double %s, metadata !"round.upward", metadata !"fpexcept.strict")
+  ; CHECK: llvm.intr.experimental.constrained.fptrunc %{{.*}} downward ignore : f64 to f32
+  %4 = call float @llvm.experimental.constrained.fptrunc.f32.f64(double %s, metadata !"round.downward", metadata !"fpexcept.ignore")
+  ; CHECK: llvm.intr.experimental.constrained.fptrunc %{{.*}} tonearestaway ignore : f64 to f32
+  %5 = call float @llvm.experimental.constrained.fptrunc.f32.f64(double %s, metadata !"round.tonearestaway", metadata !"fpexcept.ignore")
+  ; CHECK: llvm.intr.experimental.constrained.fptrunc %{{.*}} tonearestaway ignore : vector<4xf64> to vector<4xf16>
+  %6 = call <4 x half> @llvm.experimental.constrained.fptrunc.v4f16.v4f64(<4 x double> %v, metadata !"round.tonearestaway", metadata !"fpexcept.ignore")
+  ret void
+}
+
 declare float @llvm.fmuladd.f32(float, float, float)
 declare <8 x float> @llvm.fmuladd.v8f32(<8 x float>, <8 x float>, <8 x float>)
 declare float @llvm.fma.f32(float, float, float)
@@ -1120,3 +1137,5 @@ declare void @llvm.assume(i1)
 declare float @llvm.ssa.copy.f32(float returned)
 declare <vscale x 4 x float> @llvm.vector.insert.nxv4f32.v4f32(<vscale x 4 x float>, <4 x float>, i64)
 declare <4 x float> @llvm.vector.extract.v4f32.nxv4f32(<vscale x 4 x float>, i64)
+declare <4 x half> @llvm.experimental.constrained.fptrunc.v4f16.v4f64(<4 x double>, metadata, metadata)
+declare float @llvm.experimental.constrained.fptrunc.f32.f64(double, metadata, metadata)

mlir/test/Target/LLVMIR/llvmir-intrinsics.mlir

@@ -964,6 +964,35 @@ llvm.func @ssa_copy(%arg: f32) -> f32 {
   llvm.return %0 : f32
 }
 
+// CHECK-LABEL: @experimental_constrained_fptrunc
+llvm.func @experimental_constrained_fptrunc(%s: f64, %v: vector<4xf32>) {
+  // CHECK: call float @llvm.experimental.constrained.fptrunc.f32.f64(
+  // CHECK: metadata !"round.towardzero"
+  // CHECK: metadata !"fpexcept.ignore"
+  %0 = llvm.intr.experimental.constrained.fptrunc %s towardzero ignore : f64 to f32
+  // CHECK: call float @llvm.experimental.constrained.fptrunc.f32.f64(
+  // CHECK: metadata !"round.tonearest"
+  // CHECK: metadata !"fpexcept.maytrap"
+  %1 = llvm.intr.experimental.constrained.fptrunc %s tonearest maytrap : f64 to f32
+  // CHECK: call float @llvm.experimental.constrained.fptrunc.f32.f64(
+  // CHECK: metadata !"round.upward"
+  // CHECK: metadata !"fpexcept.strict"
+  %2 = llvm.intr.experimental.constrained.fptrunc %s upward strict : f64 to f32
+  // CHECK: call float @llvm.experimental.constrained.fptrunc.f32.f64(
+  // CHECK: metadata !"round.downward"
+  // CHECK: metadata !"fpexcept.ignore"
+  %3 = llvm.intr.experimental.constrained.fptrunc %s downward ignore : f64 to f32
+  // CHECK: call float @llvm.experimental.constrained.fptrunc.f32.f64(
+  // CHECK: metadata !"round.tonearestaway"
+  // CHECK: metadata !"fpexcept.ignore"
+  %4 = llvm.intr.experimental.constrained.fptrunc %s tonearestaway ignore : f64 to f32
+  // CHECK: call <4 x half> @llvm.experimental.constrained.fptrunc.v4f16.v4f32(
+  // CHECK: metadata !"round.upward"
+  // CHECK: metadata !"fpexcept.strict"
+  %5 = llvm.intr.experimental.constrained.fptrunc %v upward strict : vector<4xf32> to vector<4xf16>
+  llvm.return
+}
+
 // Check that intrinsics are declared with appropriate types.
 // CHECK-DAG: declare float @llvm.fma.f32(float, float, float)
 // CHECK-DAG: declare <8 x float> @llvm.fma.v8f32(<8 x float>, <8 x float>, <8 x float>) #0
@@ -1126,3 +1155,5 @@ llvm.func @ssa_copy(%arg: f32) -> f32 {
 // CHECK-DAG: declare ptr addrspace(1) @llvm.stacksave.p1()
 // CHECK-DAG: declare void @llvm.stackrestore.p0(ptr)
 // CHECK-DAG: declare void @llvm.stackrestore.p1(ptr addrspace(1))
+// CHECK-DAG: declare float @llvm.experimental.constrained.fptrunc.f32.f64(double, metadata, metadata)
+// CHECK-DAG: declare <4 x half> @llvm.experimental.constrained.fptrunc.v4f16.v4f32(<4 x float>, metadata, metadata)

mlir/tools/mlir-tblgen/LLVMIRConversionGen.cpp

@@ -272,6 +272,10 @@ static LogicalResult emitOneMLIRBuilder(const Record &record, raw_ostream &os,
       bs << "moduleImport.matchLocalVariableAttr";
     } else if (name == "_label_attr") {
       bs << "moduleImport.matchLabelAttr";
+    } else if (name == "_fpExceptionBehavior_attr") {
+      bs << "moduleImport.matchFPExceptionBehaviorAttr";
+    } else if (name == "_roundingMode_attr") {
+      bs << "moduleImport.matchRoundingModeAttr";
     } else if (name == "_resultType") {
       bs << "moduleImport.convertType(inst->getType())";
     } else if (name == "_location") {