[libc][math][c23] Implement canonicalize functions

[Sh0g0-1758]

Fixes: #85286

[Sh0g0-1758]

@lntue, creating this to confirm whether or not this is moving in the right direction. I am little confused in the implementation of the long double type. I will add tests once you confirm about the implementation of others. As you said :

float, double, and float128: these are IEEE 754 standard formats, most of the inputs are already canonical, with signaling NaNs are the only exception.

So I figured that they are already in canonical format and so I just need to assign *x to *cx.

[jcranmer-intel]

Some drive-by comments:

• canonicalize(SNAN) needs to raise the INVALID floating point exception
• it's a should, not a must requirement, but the corresponding quiet NaN should have the same payload as the signaling NaN (i.e., only the bit indicating whether or not the NaN is an sNaN should change)
• x86 and PPC long double types have non-canonical values that need to be handled. I don't know enough about the PPC long double type to give guidance here, but the x86 long double type considers all non-canonical values to be invalid, and will generally refuse to operate on them in the first place--glibc in this case returns 1 and does nothing.

[lntue]

Some drive-by comments:

• canonicalize(SNAN) needs to raise the INVALID floating point exception
• it's a should, not a must requirement, but the corresponding quiet NaN should have the same payload as the signaling NaN (i.e., only the bit indicating whether or not the NaN is an sNaN should change)
• x86 and PPC long double types have non-canonical values that need to be handled. I don't know enough about the PPC long double type to give guidance here, but the x86 long double type considers all non-canonical values to be invalid, and will generally refuse to operate on them in the first place--glibc in this case returns 1 and does nothing.

Thanks @jcranmer-intel ! I totally forgot to include error handling section F.10.8.7 (N3096) of canonicalize functions to the issue's description.

About the PPC long double, we will skip it for now and focusing on x86-64, arm, and riscv, which we have the hardware to test directly. Hopefully, someone with PPC knowledge and hardware could help us to fill it in later.

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[Sh0g0-1758]

Ah I got what was causing the error in the tests. Fixed them.

[lntue]

Thanks for the patch!

[nickdesaulniers]

This breaks the arm32 build in postsubmit, PTAL.
https://lab.llvm.org/buildbot/#/builders/229/builds/24394

In file included from /llvm/libc_worker/worker/libc-arm32-debian/libc-arm32-debian-dbg/llvm-project/libc/src/math/generic/sinf.cpp:11:
/llvm/libc_worker/worker/libc-arm32-debian/libc-arm32-debian-dbg/llvm-project/libc/src/__support/FPUtil/BasicOperations.h:202:10: error: no viable conversion from '__llvm_libc_19_0_0_git::BigInt<128, false>' to 'bool'
    bool bit62 = ((mantissa & (1ULL << 62)) >> 62);
         ^       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/llvm/libc_worker/worker/libc-arm32-debian/libc-arm32-debian-dbg/llvm-project/libc/src/__support/UInt.h:186:56: note: explicit conversion function is not a candidate
  template <typename T> LIBC_INLINE constexpr explicit operator T() const {
                                                       ^
/llvm/libc_worker/worker/libc-arm32-debian/libc-arm32-debian-dbg/llvm-project/libc/src/__support/UInt.h:216:34: note: explicit conversion function is not a candidate
  LIBC_INLINE constexpr explicit operator bool() const { return !is_zero(); }
                                 ^

[nickdesaulniers]

I wonder why UInt::operator bool is marked explicit? Implicit conversions to bool sure are nice.

[Sh0g0-1758]

I can provide a patch but I don't have context. Can start working on it once you provide some pointers.

[michaelrj-google]

Since the issue is that the implicit conversion to bool is failing, the easiest solution is likely to add a static_cast<bool> to the statement.

[nickdesaulniers]

Since the issue is that the implicit conversion to bool is failing, the easiest solution is likely to add a static_cast<bool> to the statement.

Cant leave the buildbots red for long: #86736

[Sh0g0-1758]

Ah sorry for the delay. My OS crashed and I saw the dreaded white death screen. Had to set up my system from scratch. Thanks for the patch @nickdesaulniers.