[flang] Define the runtime API for CPU_TIME

[rovka]

Summary

CPU_TIME takes a single real scalar INTENT(OUT) argument. We can
therefore return a double and let lowering handle casting that to the
precision used for the default real kind.

[rovka]

flang/runtime/time-intrinsic.h

22

IIUC, the default real kind is usually 4, but can be set to 8 with -fdefault-real-8.

Should we have another overload for CppTypeFor<TypeCategory::Real, 4>, or can we assume that lowering can deal with downcasting if the default kind is 4?

[schweitz]

flang/runtime/time-intrinsic.h

22

Lowering handles cases which require floating-point format conversions when calling the runtime. That said, the runtime can (and often does) provide multiple interfaces, which may eliminate the necessity of such conversions.

[klausler]

flang/runtime/time-intrinsic.h

13

I would consider putting this into the header file with miscellaneous intrinsics... were it not for the problem of SYSTEM_CLOCK (16.9.186), which will also need a home, and whose API might be messy. DATE_AND_TIME will also need to go here.

22

Lowering can handle data conversions for value arguments and for function results, but reference arguments like CPU_TIME's are somewhat more problematic. CPU_TIME (16.9.57) is a subroutine, not a function. But since its sole argument is INTENT(OUT), the API could be a function returning a C double, and lowering could take care of converting and assigning the result.

23

These source position arguments are needed only for runtime APIs that can encounter errors. CPU_TIME probably can't.

[rovka]

flang/runtime/time-intrinsic.h

13

Yes, my intention was to add SYSTEM_CLOCK here as well. As you point out, DATE_AND_TIME also fits.

22

Ok, for the sake of simplicity we can go with returning a double (although that trick won't work for SYSTEM_CLOCK, where we need 3 output parameters; but that's a different patch).

23

Ok, removing them.

[rovka]

Addressed review comments.

[jeanPerier]

LGTM

flang/runtime/time-intrinsic.h

13

FYI, there is an old draft of DATE_AND_TIME in fir-dev: https://github.com/flang-compiler/f18-llvm-project/blob/fir-dev/flang/runtime/clock.h and https://github.com/flang-compiler/f18-llvm-project/blob/fir-dev/flang/runtime/clock.cpp.
It's missing support for the VALUES argument and lack windows support. I'll upstream it in these new file when your patch is in.
It is the only intrinsic for which something was done in fir-dev only.

flang/runtime/time-intrinsic.h

+//===-- runtime/time-intrinsic.h --------------------------------*- C++ -*-===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+
+// Defines the API between compiled code and the implementations of time-related
+// intrinsic subroutines in the runtime library.
+
+#ifndef FORTRAN_RUNTIME_TIME_INTRINSIC_H_
+#define FORTRAN_RUNTIME_TIME_INTRINSIC_H_
+
+#include "cpp-type.h"
+#include "entry-names.h"
+
+namespace Fortran::runtime {
+extern "C" {
+
+// Lowering may need to cast this result to match the precision of the default
+// real kind.
+double RTNAME(CpuTime)();
+
+} // extern "C"
+} // namespace Fortran::runtime
+#endif // FORTRAN_RUNTIME_TIME_INTRINSIC_H_