[flang-rt] Runtime implementation of extended intrinsic function SECNDS() #152021

eugeneepshteyn · 2025-08-04T20:02:52Z

Until the compiler part is fully hooked up via #151878, tested this using external:

external secnds
real s1, s2
s1 = secnds(0.0)
print *, "Seconds from midnight:", s1
call sleep(2)
s2 = secnds(s1)
print *, "Seconds from s1", s2
print *, "Seconds from midnight:", secnds(0.0)
end

…DS()

klausler · 2025-08-04T21:02:37Z

flang-rt/lib/runtime/command.cpp

@@ -309,6 +310,12 @@ std::int32_t RTNAME(Hostnm)(
  return status;
 }

+float RTNAME(Secnds)(float *refTime, const char *sourceFile, int line) {


Is there a risk that lowering will emit a call with a null dummy argument pointer?

Lowering, probably not. So just not have this Secnds wrapper at all? Have lowering go directly to secnds_?

If there are other examples in lowering that work in that way, sure. Otherwise, a trivial wrapper is okay.

klausler · 2025-08-04T21:04:01Z

flang-rt/lib/runtime/extensions.cpp

+float FORTRAN_PROCEDURE_NAME(secnds)(float *refTime) {
+  constexpr float FAIL_SECNDS{-1.0f}; // Failure code for this function
+  // Failure code for time functions that return std::time_t
+  constexpr time_t FAIL_TIME{std::time_t{-1}};


maybe this should be std::time_t

klausler · 2025-08-04T21:06:21Z

flang-rt/lib/runtime/extensions.cpp

+  // comes out to about 194 days. Thus, need to pick a starting point.
+  // Given the description of this function, midnight of the current
+  // day is the best starting point.
+  static time_t startingPoint{0};


not thread-safe

should be std::time_t

not thread-safe

Definitely not. Would C++ thread_local work in this context? Or would it only work for threads created by C++ runtime?

Or dodge the problem entirely by not saving midnight, just calculating it when needed.

Here's the problem, as I see it: (1) we need to keep returned values low to not overwhelm the limited precision of float and (2) the user can call secnds() days apart for the same process runtime, so can't always go back to "this day's midnight".

Hmm, I suppose I could pre-compute "midnight" when flang-rt initializes. Is there an init function suitable for this?

Please don't add overhead to every program.

If the user passes in a non-zero reference time, you have to use it without question, yes? If the user passes in a zero reference time, it means the most recent midnight, yes?

If the user passes in a non-zero reference time, you have to use it without question, yes?

I can't use it without question, because I created it in the first place, so I know its limitations.

Let's say the user initially called s1=secnds(0.0) on Day 1 at 5 p.m., so s1 is seconds between 0:00 and 17:00. Then on Day 3 the user called s2=secnds(s1) at 10 a.m. If I just blindly use s2 - s1, the result would be wrong. I need to store something to compensate for loss of information.

I think I can use C++ atomic exchange operations to make this safer.

Be sure to not create a dependence on the C++ runtime library binaries from libflang_rt.runtime.

Yeah, I'll check if new dependencies showed up.

Dependencies before and after using atomics seem to be the same:

$ ldd ./clang/22/lib/x86_64-unknown-linux-gnu/libflang_rt.runtime.so linux-vdso.so.1 (0x00007fff0eff4000) libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007f03e025d000) libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007f03e0176000) libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007f03dff4d000) /lib64/ld-linux-x86-64.so.2 (0x00007f03e1065000)

Not sure about Windows.

akuhlens · 2025-08-04T21:15:30Z

flang-rt/lib/runtime/command.cpp

@@ -309,6 +310,12 @@ std::int32_t RTNAME(Hostnm)(
  return status;
 }

+float RTNAME(Secnds)(float *refTime, const char *sourceFile, int line) {


I assume the separation of Secnds to lower case secnds is to allow you to test it in this PR. Command.h seems to be runtime functions related to the command line, so I think extenssions.cpp/h is the right place for this not command.h. Maybe add the argument checking back to the implementation in extenssions.cpp and test the function in a unit test?

I was assuming that the other entry point was to handle calls that were to the non-intrinsic external name, as in

EXTERNAL SECNDS PRINT *, SECNDS(0.) END

I was assuming that the other entry point was to handle calls that were to the non-intrinsic external name

Yes.

Regarding Runtime/Command.h, it seems to have functions not just related to command line.

Ok, if we think providing the users both the external function implementation and the intrinsic itself, another alternative would be to put the two function functions next to each other in extension.cpp. Command.cpp still seems like the wrong place for this runtime function.

Regarding Runtime/Command.h, it seems to have functions not just related to command line.

Yeah I am pretty sure I added one of those functions without anyone telling me this was the wrong place for that function. Just because other people have made the mistake doesn't mean we should keep doing it.

I don't mind moving things, I just want to make sure things are moved to the right locations. :-)

To me, all these "command" sources seemed more like "runtime commands", not necessarily command line. Here's my understanding of these files:

flang/include/flang/Optimizer/Builder/Runtime/Command.h - MLIR building routines for runtime operations ("runtime commands" perhaps?)

flang/include/flang/Runtime/command.h - was this intended for command line argument handling only or was this intended for "runtime operations/commands" callable for MLIR building routines in above Builder/Runtime/Command.h? Right now it seems to contain all _FortranA... runtime routines callable by lowering.

flang-rt/lib/runtime/command.cpp - currently it's implementation of _FortranA... runtime routines declared in flang/include/flang/Runtime/command.h above.

flang/include/flang/Runtime/extensions.h - declarations for all the extensions, suitable for external usage.

flang-rt/lib/runtime/extensions.cpp - implementations for all the extensions.

So the question is whether the wrapper _FortranA... routines for the extensions should move to extensions.cpp. It seems to me that they should be together with the other wrapper routines in flang-rt/lib/runtime/command.cpp

command.h/.cpp should only contain things related to the command-line intrinsics. It accumulated more, but certainly not all of the RTNAME(...) entry points, probably due to initial laziness and later following of poor examples. The miscellaneous stuff like HostNm really shouldn't be in there.

Ok, I shall move secnds() stull to extensions.cpp, as @akuhlens suggested.

Andre, since you had the right idea, perhaps you want to do a general clean-up patch to move things to the right places?

… time_t declarations

flang-rt/lib/runtime/extensions.cpp

eugeneepshteyn · 2025-08-05T03:39:46Z

flang-rt/lib/runtime/extensions.cpp

+  static std::atomic<std::time_t> startingPoint{TIME_UNINITIALIZED};
+  std::time_t expected{TIME_UNINITIALIZED};
+  std::time_t localStartingPoint{TIME_UNINITIALIZED};
+  if (startingPoint.compare_exchange_strong(expected, TIME_INITIALIZING)) {


Attempt to support multiple threads. Only tested single threaded for now. Will test with OpenMP threading.

eugeneepshteyn · 2025-08-05T19:55:18Z

flang-rt/lib/runtime/extensions.cpp

@@ -303,6 +304,93 @@ void FORTRAN_PROCEDURE_NAME(qsort)(int *array, int *len, int *isize,
 // PERROR(STRING)
 void RTNAME(Perror)(const char *str) { perror(str); }

+// GNU extension function SECNDS(refTime)


This is ready for review. I did some testing with multiple threads. Test program:

use omp_lib external secnds real sec_mid real results(1000) integer i, num_threads results = 0.0 !$omp parallel do do i=1, size(results) results(i) = secnds(0.0) num_threads = omp_get_num_threads() end do !$omp end parallel do print *, "Number of threads in parallel region:", num_threads sec_mid = minval(results) print *, "Seconds from midnight: min:", sec_mid, ", max:", maxval(results) call sleep(2) !$omp parallel do do i=1, size(results) results(i) = secnds(sec_mid) end do !$omp end parallel do print *, "Seconds from sec_mid: min:", minval(results), ", max:", maxval(results) print *, "Seconds from midnight:", secnds(0.0) end

Output:

$ ./a.out Number of threads in parallel region: 256 Seconds from midnight: min: 46489. , max: 46489. Seconds from sec_mid: min: 2. , max: 2. Seconds from midnight: 46491.

eugeneepshteyn · 2025-08-05T20:00:05Z

@mlir-maiden , FYI

klausler

I think that you can replace that loop with just a single compare & swap if zero. Only the first thread to ever execute it would succeed, and all later ones would fail and return the first thread's value.

flang-rt/lib/runtime/extensions.cpp

github-actions · 2025-08-05T20:19:40Z

✅ With the latest revision this PR passed the C/C++ code formatter.

eugeneepshteyn · 2025-08-05T20:24:06Z

I think that you can replace that loop with just a single compare & swap if zero. Only the first thread to ever execute it would succeed, and all later ones would fail and return the first thread's value.

The other threads still need to wait for the initialization to happen, so I think the loop is unavoidable. (In early implementation, I didn't have a loop and threads came in quicker than the time value could be computed.) Also, I decided to put even compare_exchange_strong() inside the loop, because I read somewhere that even this function can still fail under some conditions (for example, see section in https://dev.to/pauljlucas/advanced-thread-safety-in-c-3ap5).

klausler · 2025-08-05T20:34:00Z

I think that you can replace that loop with just a single compare & swap if zero. Only the first thread to ever execute it would succeed, and all later ones would fail and return the first thread's value.

The other threads still need to wait for the initialization to happen, so I think the loop is unavoidable. (In early implementation, I didn't have a loop and threads came in quicker than the time value could be computed.) Also, I decided to put even compare_exchange_strong() inside the loop, because I read somewhere that even this function can still fail under some conditions (for example, see section in https://dev.to/pauljlucas/advanced-thread-safety-in-c-3ap5).

The atomic compare & swap if zero would be the initialization, I mean. The flow would be like:

  auto UTCmidnight{atomicRead(atomicVar}};
  if (UTCmidnight == 0) { // not initialized
    // not initialized yet
    UTCmidnight = computation;
    CHECK(UTCmidnight > 0);
    auto value{atomicCompareAndSwap(atomicVar, 0, UTCmidnight)};  // only stores a value once ever
    if (oldValue != 0) {
      // lost the race, my value wasn't stored
      UTCmidnight = oldValue;
    }
  }
  // all threads have the same UTCmidnight value from this point

eugeneepshteyn · 2025-08-05T20:38:39Z

The atomic compare & swap if zero would be the initialization, I mean.

Ah, ok, at the cost of possible extra "midnight" computations. I'll try that.

flang/include/flang/Runtime/command.h

klausler · 2025-08-06T16:45:58Z

flang-rt/lib/runtime/extensions.cpp

+  // Initialize startingPoint if we haven't initialized it yet or
+  // if we were passed 0.0f, which indicates to compute seconds from
+  // current day's midnight.
+  if (localStartingPoint == TIME_UNINITIALIZED || *refTime < 0.5f) {


Is SECNDS() documented to look for < 0.5 or for == 0.?

In practice this would be whole number of seconds represented as float.

I'm always cautious about exact real numbers comparisons. Will == 0.0f always work here?

You don't need the f suffix, and the comparison will be true for both 0. and -0. inputs. If that's what was documented, it's best to stick to it.

klausler · 2025-08-06T16:46:42Z

flang-rt/lib/runtime/extensions.cpp

@@ -303,6 +304,75 @@ void FORTRAN_PROCEDURE_NAME(qsort)(int *array, int *len, int *isize,
 // PERROR(STRING)
 void RTNAME(Perror)(const char *str) { perror(str); }

+// GNU extension function SECNDS(refTime)
+float FORTRAN_PROCEDURE_NAME(secnds)(float *refTime) {


It would be a good idea to write this as a template now so that it doesn't need to converted later when somebody gets around to implementing SECNDSD.

eugeneepshteyn added 7 commits August 4, 2025 16:00

[flang-rt] Runtime implementation of extended intrinsic function SECN…

68a1a81

…DS()

clang-format

716f82d

Fixed a typo

8bbe4eb

Fixed init form

5f671e7

The failure code should be negative

9c53878

Introduce FAIL_TIME

80b61ed

Merge branch 'llvm:main' into secnds-runtime

26b1265

eugeneepshteyn requested review from klausler and akuhlens August 4, 2025 20:56

eugeneepshteyn mentioned this pull request Aug 4, 2025

[flang] Register and lower SECNDS (stubbed implementation) #151878

Open

eugeneepshteyn marked this pull request as ready for review August 4, 2025 20:58

llvmbot added the flang Flang issues not falling into any other category label Aug 4, 2025

klausler reviewed Aug 4, 2025

View reviewed changes

akuhlens requested changes Aug 4, 2025

View reviewed changes

eugeneepshteyn added 4 commits August 4, 2025 23:05

Code review: moved RTNAME(Secnds) to extensions.cpp. Added missing to…

78a8d07

… time_t declarations

Attempt at using atomic operations in secnds_() to make it reentrant

2f3a274

clang-format

9940825

Fixed init

7268c8c

eugeneepshteyn commented Aug 5, 2025

View reviewed changes

eugeneepshteyn added 4 commits August 5, 2025 06:22

Merge branch 'llvm:main' into secnds-runtime

812760f

Reset startingPoint to 'uninitialized' on failure

205818e

Changed currentStartingPoint initialization to use loop in all cases

5945264

clang-format

5cdb74d

eugeneepshteyn commented Aug 5, 2025

View reviewed changes

eugeneepshteyn requested review from akuhlens and klausler August 5, 2025 19:55

klausler reviewed Aug 5, 2025

View reviewed changes

flang-rt/lib/runtime/extensions.cpp Outdated Show resolved Hide resolved

flang-rt/lib/runtime/extensions.cpp Outdated Show resolved Hide resolved

Code review feedback

599713e

clang-format

4ae3cd3

akuhlens reviewed Aug 5, 2025

View reviewed changes

flang/include/flang/Runtime/command.h Outdated Show resolved Hide resolved

Moved RTNAME(Secnds) declaration to extensions.h

fdd310a

akuhlens self-requested a review August 6, 2025 00:42

eugeneepshteyn added 3 commits August 5, 2025 20:54

Simplified the implementation based on Peter's suggestion

1cf3a46

clang-format

aa3d3d4

Merge branch 'llvm:main' into secnds-runtime

0603b56

klausler reviewed Aug 6, 2025

View reviewed changes

[flang-rt] Runtime implementation of extended intrinsic function SECNDS() #152021

Are you sure you want to change the base?

[flang-rt] Runtime implementation of extended intrinsic function SECNDS() #152021

Conversation

eugeneepshteyn commented Aug 4, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

eugeneepshteyn Aug 4, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

eugeneepshteyn commented Aug 5, 2025

Uh oh!

klausler left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

github-actions bot commented Aug 5, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

eugeneepshteyn commented Aug 5, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

klausler commented Aug 5, 2025

Uh oh!

eugeneepshteyn commented Aug 5, 2025

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

eugeneepshteyn commented Aug 4, 2025 •

edited

Loading

eugeneepshteyn Aug 4, 2025 •

edited

Loading

github-actions bot commented Aug 5, 2025 •

edited

Loading

eugeneepshteyn commented Aug 5, 2025 •

edited

Loading