[Flang] PURE procedures incorrectly accept calls to impure intrinsics

[mlir-maiden]

It seems that Flang does not enforce restrictions on usage of impure intrinsics inside PURE procedures. Based on my tests, programs that gfortran correctly rejects as non-conforming compile and run successfully in Flang. This is a standards-conformance gap.

Example
Using CPU_TIME (impure subroutine):

program get_cpu_time
  print *, measure_cpu_time()
contains
  pure function measure_cpu_time() result(elapsed_time)
    real :: elapsed_time
    real :: time_start, time_end

    call cpu_time(time_start)   ! impure intrinsic inside PURE
    call cpu_time(time_end)     

    elapsed_time = time_end - time_start
  end function measure_cpu_time
end program get_cpu_time

Result in gfortran
Error: Subroutine call to intrinsic 'cpu_time' at (1) is not PURE

Result in flang
Compiles and runs, prints a time value.

Other intrinsics tested
The same behavior was observed with:

• SYSTEM_CLOCK
• DATE_AND_TIME
• ETIME (subroutine + function forms)
• SECNDS

It seems that Flang's checks for PURE procedures do not currently consult intrinsic classifications, allowing impure intrinsics to slip through.

Proposed Fix
Update semantic checks for PURE procedures to reject calls to intrinsics classified as impure.
I am a beginner contributor, but I would be happy to help with a patch once I have guidance on where in the semantics pipeline this check should be added. Thank you

[eugeneepshteyn]

ifx also rejects it:

$ ifx -c impure.f90 
impure.f90(8): error #7137: Any procedure referenced in a PURE procedure, including one referenced via a defined operation or assignment, must have an explicit interface and be declared PURE.   [CPU_TIME]
    call cpu_time(time_start)   ! impure intrinsic inside PURE
---------^
impure.f90(9): error #7137: Any procedure referenced in a PURE procedure, including one referenced via a defined operation or assignment, must have an explicit interface and be declared PURE.   [CPU_TIME]
    call cpu_time(time_end)     
---------^
compilation aborted for impure.f90 (code 1)

@mlir-maiden , do you plan to work on this issue?

[mlir-maiden]

@eugeneepshteyn , I definitely want to take this on. I am very new to Flang, so any pointers in the right direction would be much appreciated. Thank you

[klausler]

It's a useful extension and it doesn't violate the purpose having PURE procedures. Looking at the clocks doesn't risk modifying program-visible state. Excluding these timing utilities from being used in a PURE procedure would just make it more difficult to analyze their performance. Perhaps these calls deserve an optional portability warning under -pedantic, but they should not be errors.

(Other impure subroutines that aren't timing-related, like SYSTEM, should be excluded from PURE procedures, if you want to do that.)

[sscalpone]

Is this extension documented in Extensions.md?

[mlir-maiden]

@sscalpone Thanks, Steve. I checked flang/docs/Extensions.md and didn’t find this behavior documented. The only PURE-related entry there is about allowing dummy arguments that aren’t INTENT(IN) or VALUE. There are no entries for CPU_TIME, SYSTEM_CLOCK, DATE_AND_TIME, etc. with respect to PURE, so it looks like this isn’t a documented extension yet.

From Peter’s comment, I understand the intended handling is two-fold:
• Timing intrinsics allowed in PURE (with an optional -pedantic warning),
• Other impure intrinsics excluded from PURE with the proper semantics checks.

I’d be glad to take this on, and update Extensions.md accordingly.

[eugeneepshteyn]

Ok, @mlir-maiden , this issue is yours.

[llvmbot]

@llvm/issue-subscribers-flang-frontend

Author: Šárka Holendová (mlir-maiden)

It seems that Flang does not enforce restrictions on usage of impure intrinsics inside PURE procedures. Based on my tests, programs that gfortran correctly rejects as non-conforming compile and run successfully in Flang. This is a standards-conformance gap.

Example
Using CPU_TIME (impure subroutine):

program get_cpu_time
  print *, measure_cpu_time()
contains
  pure function measure_cpu_time() result(elapsed_time)
    real :: elapsed_time
    real :: time_start, time_end

    call cpu_time(time_start)   ! impure intrinsic inside PURE
    call cpu_time(time_end)     

    elapsed_time = time_end - time_start
  end function measure_cpu_time
end program get_cpu_time

Result in gfortran
Error: Subroutine call to intrinsic 'cpu_time' at (1) is not PURE

Result in flang
Compiles and runs, prints a time value.

Other intrinsics tested
The same behavior was observed with:

• SYSTEM_CLOCK
• DATE_AND_TIME
• ETIME (subroutine + function forms)
• SECNDS

It seems that Flang's checks for PURE procedures do not currently consult intrinsic classifications, allowing impure intrinsics to slip through.

Proposed Fix
Update semantic checks for PURE procedures to reject calls to intrinsics classified as impure.
I am a beginner contributor, but I would be happy to help with a patch once I have guidance on where in the semantics pipeline this check should be added. Thank you

[klausler]

Are you still planning to fix this bug soon? If not, please reassign it to me.

[mlir-maiden]

Are you still planning to fix this bug soon? If not, please reassign it to me.

I would like to work on this but cannot turn it around quickly. If time is of the essence, I don’t want to be a bottleneck, so I’ll assign it to you. I’ll follow the patch to learn from it. Thanks for picking it up.