libzigc/math: Implement rintl, lrintl #31791

It's a fairly straightforward port of musl's rintl, like rint and rintf were. libc-test tests for rintl are uncommented since they're now passing, and, as always, the results are attached to the relevant implementation commits.

I've also covered special cases for rint with tests, and broke down the current rint and modf test declarations into multiple -- so each libc function get its own test declaration at the very least.

Contributes to #30978

@mihael wrote in #31791 (comment):

But this is something I'd like to get back to. I'm playing around with an idea of a packed struct that holds a bit representation of a float and offers a nice API for these bit manipulation/checking operations. Ideally, it would also support all the float types, not just the long double ones. That would come in handy when dealing with musl ports and making them a little easier to read and work with. Should it work, it would supersede the long double namespace.

Though the FloatRepr is on track of what I was thinking of, just with more abstraction than I was envisioning, but that shouldn't be a detractor if there's no performance penalty. We could lean on using FloatRepr in more places and a couple of useful helper methods.

Wasn't aware it was blocked, but I'm running into some problems during the test-modules step in Debug mode after rebasing on master. When I run the libc unit tests:

./build/stage3/bin/zig build test-modules test-c-abi -Dtest-target-filter=musl -Dtest-target-filter=windows-gnu -fqemu -fwasmtime --summary all

it falls apart across all targets in Debug mode, e.g.:

• Wrong values for actual results
• Special cases also mangled (0 instead of inf, etc.)
• s390x crashes
• Crashes for WASM:

test-modules
└─ test-libc
 └─ run test libc-wasm32-wasi-musl-lime1-Debug-libc 2 pass, 1 crash (3 total)
error: 'c.math.test.rintl.normal' exited with code 134 with stderr:
 Error: failed to run main module `./.zig-cache/o/c8ad29b82a33211ea61b45233946f344/test.wasm`
 
 Caused by:
 0: failed to invoke command default
 1: error while executing at wasm backtrace:
 0: 0x5a91a - test.wasm!.Lc.math.rintl_bitcast_invalid
 1: 0x59a0d - test.wasm!c.math.testRintNormal__anon_25472
 2: 0x5a922 - test.wasm!c.math.test.rintl.normal
 3: 0x5ae5 - test.wasm!test_runner.mainServer
 4: 0x3a31 - test.wasm!test_runner.main
 5: 0x34a9 - test.wasm!start.mainWithoutEnv
 6: 0xca4 - test.wasm!__main_void
 7: 0xc07 - test.wasm!_start
 note: using the `WASMTIME_BACKTRACE_DETAILS=1` environment variable may show more debugging information
 2: wasm trap: wasm `unreachable` instruction executed
failed command: wasmtime --dir=. -Sinherit-env ./.zig-cache/o/c8ad29b82a33211ea61b45233946f344/test.wasm --cache-dir=./.zig-cache --seed=0xac1f0997 --listen=-

etc.

Yet ReleaseFast and ReleaseSafe are fine.

That's why I added the Debug skips for rintl tests:

if(builtin.mode==std.builtin.OptimizeMode.Debug)returnerror.SkipZigTest;// TODO

(On a second thought, it'd be better add a filter for the test-modules step to the skip -- should be possible I believe. But I'd have to additionally test that & I wanted to share this before logging off for the day -- more precisely, night :)).

Commenting them out will result in errors when running the above command.
libc-aarch64-linux-musl-generic-Debug-libc, libc-aarch64_be-linux-musl-generic-Debug-libc, and libc-x86_64-linux-musl-x86_64-Debug-libc-llvm-no-lld seem fine even without the skips.

I had investigated, but I'm none the wiser. At first, I thought maybe it could be an issue that I use const toint = 1.0 / math.floatEps(c_longdouble);, the divisor (LDBL EPSILON) would be 0.00000000000000000000000000000000019259299443872358530559779425849273 for f128, whereas musl uses the 80-bit epsilon (1.0842021724855044340e-19) both for f80 and f128 on most architectures (ex., source_code for LDBL_EPSILON, source code for rintl impl). But the issues still persisted.

All tests pass for libc-test suite:

$ ./build/stage3/bin/zig build test-libc-nsz -Dlibc-test-path=<PATH-TO-LIBC-TEST-SUITE> -fqemu -fwasmtime --summary line
Build Summary: 49121/49121 steps succeeded

Other than resolving the conflicts, I also changed the stuff that is deprecated in 0.16 (@intFromFloat, std.meta.Int).
If anyone has any clue why Debug falls apart like that, I'd appreciate it.

This should be it. I've skipped the rintl tests on PowerPC and referenced the issue (#30976) in the comment. I'm also skipping, but documenting with TODO, the two test cases (floats close to -n.5 rounding to -n.5) that fail on NetBSD. Since libzigc is not being built for NetBSD, it would seem it's a bug in their native libc -- I guess.