[flang] Lower REAL(16) MODULO to Float128Math library call. (#85322)

I did not test it through in #85005, and my assumption was wrong:
arith::RemFOp might be lowered to an fmodf128() call that does not
exist everywhere.

Author: vzakhari

flang/include/flang/Optimizer/Builder/Runtime/Numeric.h

@@ -30,6 +30,10 @@ mlir::Value genFraction(fir::FirOpBuilder &builder, mlir::Location loc,
 mlir::Value genMod(fir::FirOpBuilder &builder, mlir::Location loc,
                    mlir::Value a, mlir::Value p);
 
+/// Generate call to Modulo intrinsic runtime routine.
+mlir::Value genModulo(fir::FirOpBuilder &builder, mlir::Location loc,
+                      mlir::Value a, mlir::Value p);
+
 /// Generate call to Nearest intrinsic runtime routine.
 mlir::Value genNearest(fir::FirOpBuilder &builder, mlir::Location loc,
                        mlir::Value x, mlir::Value s);

flang/lib/Optimizer/Builder/IntrinsicCall.cpp

@@ -5259,6 +5259,13 @@ mlir::Value IntrinsicLibrary::genModulo(mlir::Type resultType,
                                                  remainder);
   }
 
+  // F128 arith::RemFOp may be lowered to a runtime call that may be unsupported
+  // on the target, so generate a call to Fortran Runtime's ModuloReal16.
+  if (resultType == mlir::FloatType::getF128(builder.getContext()))
+    return builder.createConvert(
+        loc, resultType,
+        fir::runtime::genModulo(builder, loc, args[0], args[1]));
+
   auto remainder = builder.create<mlir::arith::RemFOp>(loc, args[0], args[1]);
   mlir::Value zero = builder.createRealZeroConstant(loc, remainder.getType());
   auto remainderIsNotZero = builder.create<mlir::arith::CmpFOp>(

flang/lib/Optimizer/Builder/Runtime/Numeric.cpp

@@ -118,6 +118,20 @@ struct ForcedMod16 {
   }
 };
 
+/// Placeholder for real*16 version of Modulo Intrinsic
+struct ForcedModulo16 {
+  static constexpr const char *name = ExpandAndQuoteKey(RTNAME(ModuloReal16));
+  static constexpr fir::runtime::FuncTypeBuilderFunc getTypeModel() {
+    return [](mlir::MLIRContext *ctx) {
+      auto fltTy = mlir::FloatType::getF128(ctx);
+      auto strTy = fir::ReferenceType::get(mlir::IntegerType::get(ctx, 8));
+      auto intTy = mlir::IntegerType::get(ctx, 8 * sizeof(int));
+      return mlir::FunctionType::get(ctx, {fltTy, fltTy, strTy, intTy},
+                                     {fltTy});
+    };
+  }
+};
+
 /// Placeholder for real*10 version of Nearest Intrinsic
 struct ForcedNearest10 {
   static constexpr const char *name = ExpandAndQuoteKey(RTNAME(Nearest10));
@@ -323,6 +337,33 @@ mlir::Value fir::runtime::genMod(fir::FirOpBuilder &builder, mlir::Location loc,
   return builder.create<fir::CallOp>(loc, func, args).getResult(0);
 }
 
+/// Generate call to Modulo intrinsic runtime routine.
+mlir::Value fir::runtime::genModulo(fir::FirOpBuilder &builder,
+                                    mlir::Location loc, mlir::Value a,
+                                    mlir::Value p) {
+  mlir::func::FuncOp func;
+  mlir::Type fltTy = a.getType();
+
+  if (fltTy != p.getType())
+    fir::emitFatalError(loc, "arguments type mismatch in MOD");
+
+  // MODULO is lowered into math operations in intrinsics lowering,
+  // so genModulo() should only be used for F128 data type now.
+  if (fltTy.isF128())
+    func = fir::runtime::getRuntimeFunc<ForcedModulo16>(loc, builder);
+  else
+    fir::intrinsicTypeTODO(builder, fltTy, loc, "MODULO");
+
+  auto funcTy = func.getFunctionType();
+  auto sourceFile = fir::factory::locationToFilename(builder, loc);
+  auto sourceLine =
+      fir::factory::locationToLineNo(builder, loc, funcTy.getInput(3));
+  auto args = fir::runtime::createArguments(builder, loc, funcTy, a, p,
+                                            sourceFile, sourceLine);
+
+  return builder.create<fir::CallOp>(loc, func, args).getResult(0);
+}
+
 /// Generate call to Nearest intrinsic runtime routine.
 mlir::Value fir::runtime::genNearest(fir::FirOpBuilder &builder,
                                      mlir::Location loc, mlir::Value x,

flang/test/Lower/Intrinsics/modulo.f90

@@ -40,17 +40,6 @@ subroutine modulo_testi(r, a, p)
 ! CHECK-SAME: %[[arg0:.*]]: !fir.ref<f128>{{.*}}, %[[arg1:.*]]: !fir.ref<f128>{{.*}}, %[[arg2:.*]]: !fir.ref<f128>{{.*}}) {
 subroutine modulo_testr16(r, a, p)
   real(16) :: r, a, p
-  ! CHECK-DAG: %[[a:.*]] = fir.load %[[arg1]] : !fir.ref<f128>
-  ! CHECK-DAG: %[[p:.*]] = fir.load %[[arg2]] : !fir.ref<f128>
-  ! CHECK-DAG: %[[rem:.*]] = arith.remf %[[a]], %[[p]] {{.*}}: f128
-  ! CHECK-DAG: %[[zero:.*]] = arith.constant 0.000000e+00 : f128
-  ! CHECK-DAG: %[[remNotZero:.*]] = arith.cmpf une, %[[rem]], %[[zero]] {{.*}} : f128
-  ! CHECK-DAG: %[[aNeg:.*]] = arith.cmpf olt, %[[a]], %[[zero]] {{.*}} : f128
-  ! CHECK-DAG: %[[pNeg:.*]] = arith.cmpf olt, %[[p]], %[[zero]] {{.*}} : f128
-  ! CHECK-DAG: %[[signDifferent:.*]] = arith.xori %[[aNeg]], %[[pNeg]] : i1
-  ! CHECK-DAG: %[[mustAddP:.*]] = arith.andi %[[remNotZero]], %[[signDifferent]] : i1
-  ! CHECK-DAG: %[[remPlusP:.*]] = arith.addf %[[rem]], %[[p]] {{.*}}: f128
-  ! CHECK: %[[res:.*]] = arith.select %[[mustAddP]], %[[remPlusP]], %[[rem]] : f128
-  ! CHECK: fir.store %[[res]] to %[[arg0]] : !fir.ref<f128>
+  ! CHECK: fir.call @_FortranAModuloReal16({{.*}}){{.*}}: (f128, f128, !fir.ref<i8>, i32) -> f128
   r = modulo(a, p)
 end subroutine