Alternatives to ieee_value for creating exceptional values about toml-f

[awvwgk]

While a somewhat unsatisfying solution, it seems to pass the TOML testsuite. Since those branches are only relevant for reading exceptional values they are possibly not performance critical, so this should be an okay-ish solution for the time being.

[awvwgk]

I know, I only removed cases in the library and not the tests.

[awvwgk]

The ieee_value function is used to create values for NaN, Inf and -Inf at

toml-f/src/tomlf/de/lexer.f90

Lines 1297 to 1309 in 6290b0b

	if (match(lexer, first, "n")) then
	val = ieee_value(val, ieee_quiet_nan)
	return
	end if
	if (match(lexer, first, "i")) then
	if (match(lexer, token%first, char_kind%minus)) then
	val = ieee_value(val, ieee_negative_inf)
	else
	val = ieee_value(val, ieee_positive_inf)
	end if
	return
	end if

If there is a standard conforming way to implement these without ieee_value I'm happy to use those. Note that the TOML standard requires a compliant parser to provide support for IEEE 754 binary64 values (see here).

[awvwgk]

One quick hack would be to replace the values with BOZ literals. But this might actually be less portable, since it does not account for endianess.

 if (match(lexer, first, "n")) then 
    val = real(z'7FF8000000000001', tfr)
    return 
 end if 
  
 if (match(lexer, first, "i")) then 
    if (match(lexer, token%first, char_kind%minus)) then 
       val = real(z'FFF0000000000000', tfr)
    else 
       val = real(z'7FF0000000000000', tfr)
    end if 
    return 
 end if 

[barracuda156]

@awvwgk Unfortunately I cannot come up with any suggestions as long as Fortran goes. (C++ has implementations for those.)

Using code which does not account for endianness would be a step backward IMO. That will break more platforms than we have broken now :)
To be clear, I am actually looking for a generally optimal solution here, not for a hack to make it work for me. For me it works, I have an experimental fix for gfortran which allows building ieee_arithmetic module. (The error I quoted is from Rosetta system; the same version of fpm builds fine on my PowerMac with kinda-fixed gfortran.)

[awvwgk]

At least gfortran will create errors in case one tries to use the tricks described in the C++ docs to create exceptional values:

❯ cat exceptional.f90
program exceptional
  implicit none

  print '(g0)', &
    acos(2.0), &
    log(-1.0), &
    sqrt(-1.0)

  print '(g0)', &
    huge(1.0)*2
end program exceptional
❯ gfortran exceptional.f90&& ./a.out
exceptional.f90:5:9:

    5 |     acos(2.0), &
      |         1
Error: Argument of ACOS at (1) must be between -1 and 1
exceptional.f90:6:8:

    6 |     log(-1.0), &
      |        1
Error: Argument of LOG at (1) cannot be less than or equal to zero
exceptional.f90:7:9:

    7 |     sqrt(-1.0)
      |         1
Error: Argument of SQRT at (1) has a negative value
exceptional.f90:10:8:

   10 |     huge(1.0)*2
      |        1
Error: Arithmetic overflow at (1)

[awvwgk]

Currently I cannot come up with a better solution other than ieee_value which is the standard conforming way to obtain those values (in Fortran there is usually one way and only one way to archive most things).

[awvwgk]

The only more general fix I can come up with is an internal read, which I am actually trying to avoid in this piece of the code...

 if (match(lexer, first, "n")) then 
    read("NaN", *) val
    return 
 end if 
  
 if (match(lexer, first, "i")) then 
    read("Inf", *) val
    if (match(lexer, token%first, char_kind%minus)) then 
       val = -val
    end if 
    return 
 end if 

[barracuda156]

The only more general fix I can come up with is an internal read, which I am actually trying to avoid in this piece of the code...

 if (match(lexer, first, "n")) then 
    read("NaN", *) val
    return 
 end if 
  
 if (match(lexer, first, "i")) then 
    read("Inf", *) val
    if (match(lexer, token%first, char_kind%minus)) then 
       val = -val
    end if 
    return 
 end if 

Is the fallback code feasible or too complicated? Some macro that gonna determine if ieee_arithmetic is supported by the compiler, if not, then use suboptimal, but still working code.

[barracuda156]

@awvwgk Thank you!

[barracuda156]

@awvwgk Turned out, there are more instances:

/opt/local/var/macports/build/_opt_PPCRosettaPorts_fortran_toml-f/toml-f/work/build/_deps/test-drive-src/test/test_select.F90:15:20:

   15 |   use, intrinsic :: ieee_arithmetic, only : ieee_value, ieee_quiet_nan
      |                    1
Fatal Error: Cannot find an intrinsic module named 'ieee_arithmetic' at (1)
compilation terminated.
/opt/local/var/macports/build/_opt_PPCRosettaPorts_fortran_toml-f/toml-f/work/build/_deps/test-drive-src/test/test_check.F90:25:20:

   25 |   use, intrinsic :: ieee_arithmetic, only : ieee_value, ieee_quiet_nan
      |                    1

[barracuda156]

I know, I only removed cases in the library and not the tests.

Got it, thanks. Maybe add an option to build without tests then? Now it is unconditional:

# add the testsuite
enable_testing()
set(fpm-toml "${PROJECT_SOURCE_DIR}/fpm.toml")
add_subdirectory("test")