[Fortran] Add floating point to integer conversion test

[ashermancinelli]

Semantics for floating point conversions to integer types was changed in flang to use the saturated floating point intrinsics to more closely match the standard semantics in llvm-project#130686 . This tests conversions from tiny, huge, and inf to various integer types.

[jeanPerier]

Thanks for adding this test here! Works for me.

[ashermancinelli]

A couple notes:

• We can remove unsigned if folks use this test suite with compilers that don't support that flag (or disable this test under some flag)
• I wouldn't mind adding real(kind=16) support if we have a way to tell if it's supported with the current compiler

[tarunprabhu]

* We can remove unsigned if folks use this test suite with compilers that don't support that flag (or disable this test under some flag)

I don't think the llvm-test-suite is really intended to be a general-purpose compiler test suite. Unless flang has a way to disable support for unsigned, I wouldn't worry about it.

* I wouldn't mind adding real(kind=16) support if we have a way to tell if it's supported with the current compiler

I was thinking that myself but I don't know how to check for this off the top of my head. Maybe @jeanPerier knows, or knows of someone who does.

[tblah]

I was thinking that myself but I don't know how to check for this off the top of my head. Maybe @jeanPerier knows, or knows of someone who does.

! REQUIRES: flang-supports-f128-math

EDIT: Apologies this would actually only work for a lit test.

I think for unit tests you would need to duplicate the CMake logic for f128 support. You can see examples here https://github.com/llvm/llvm-project/pull/106957/files

[ashermancinelli]

Thanks @tblah! If this test suite is not trying to be portable, maybe we could use the FLANG_SUPPORT_R16 macro?

[tarunprabhu]

This worked for me on Linux on x86. Thanks.

[ashermancinelli]

I split the test into two, one of which uses real(16) and is conditionally enabled.

Also, since this is just for LLVM I added a quick check for NaN -> int conversions.

Thanks for the reviews! Please LMK if there's anything else y'all would like changed.

[tblah]

Thanks for the update.

This correctly enabled r16 support and passed on my aarch64 system.

[jeanPerier]

Thanks

[tarunprabhu]

Does fp_convert_r16.f90 need to do the same thing as fp_convert.f90 in addition to kind=16? It looks like fp_convert.f90 is always built and run. Could we do just kind=16 in the other file then or am I missing something obvious>

[ashermancinelli]

Does fp_convert_r16.f90 need to do the same thing as fp_convert.f90 in addition to kind=16?

The fp_convert_r16 test converts to some of the same types, but from real16 and also to int16 and unsigned16 - there should be no conversions that happen in both test cases. Is that what you were asking?

[tarunprabhu]

Does fp_convert_r16.f90 need to do the same thing as fp_convert.f90 in addition to kind=16?

The fp_convert_r16 test converts to some of the same types, but from real16 and also to int16 and unsigned16 - there should be no conversions that happen in both test cases. Is that what you were asking?

Ah, sorry. I didn't notice that value in do_conversion is real(8) in one and real(16) in the other.

Thanks for clarifying. LGTM