tests: use identical(NA_real_+0, NA_real_) to skip tests that depend on NA payload propagation

[aitap]

RISC-V chips do not propagate NaN payload through arithmetic operations, so skip two tests on this architecture. If this surfaces again, we can modify test() to ignore NA / NaN differences on such architectures. Ironically, the tests in froll.Rraw that are skipped on RISC-V due to long double size ("under Valgrind") do pass on Valgrind now, but they probably won't pass on RISC-V (for the same reason) if enabled. Should I reuse the exact_NaN test instead?

@charles-plessy, this seems to work for me in a Trixie QEMU chroot. Does anything else need fixing?

Fixes: #7695

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.04%. Comparing base (76ec862) to head (d4ba469).
⚠️ Report is 1 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #7702   +/-   ##
=======================================
  Coverage   99.04%   99.04%           
=======================================
  Files          87       87           
  Lines       17037    17037           
=======================================
  Hits        16874    16874           
  Misses        163      163           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[charles-plessy]

Thanks a lot! In my understanding we do not need to fix anything else.

[MichaelChirico]

Is this truly unique to riscv64? I saw something in another thread about this platform supporting 128-bit long doubles, I think this issue is separate, right?

I also saw this might be the same as some behavior observable with {base}, could this be a behavioral skip instead? if (identical(NA+0, NaN)) something like that (sorry I didn't read very carefully)?

Basically my preference ordering here:

1. Skip on a behavior observable in base R
2. Skip on capabilities()
3. Skip on a specific Sys.info() entry

I trust your judgment on if either of 1/2 are feasible without inordinate effort. Thanks!

[aitap]

Of course, you're right! Checking whether an arithmetic operation on NA_real_ returns NA or NaN is the right answer here. Tested again with -d valgrind, inside the QEMU chroot, and on an R build configured not to use long double.