[libc] -Wnarrowing in many _Float16 tests with gcc-14

[nickdesaulniers]

Running ninja libc-math-smoke-tests with gcc-14 produces a spew of:

llvm-project/libc/test/src/math/smoke/exp2m1f16_test.cpp:74:15: error: ISO C++ does not allow converting to ‘float16’ {aka ‘_Float16’} from ‘double’ with greater conversion rank [-Werror=narrowing]
   74 |   float16 x = -12.0;
      |               ^~~~~

(in many different test files). This is going to block enabling newer gcc in presubmit, and harm our ability to test with gcc until resolved.

cc @lntue @wldfngrs @overmighty

found while trying to test #124036

[llvmbot]

@llvm/issue-subscribers-libc

Author: Nick Desaulniers (nickdesaulniers)

Running `ninja libc-math-smoke-tests` with gcc-14 produces a spew of: ``` llvm-project/libc/test/src/math/smoke/exp2m1f16_test.cpp:74:15: error: ISO C++ does not allow converting to ‘float16’ {aka ‘_Float16’} from ‘double’ with greater conversion rank [-Werror=narrowing] 74 | float16 x = -12.0; | ^~~~~ ``` (in many different test files). This is going to block enabling newer gcc in presubmit, and harm our ability to test with gcc until resolved.

cc @lntue @wldfngrs @overmighty

[nickdesaulniers]

https://developer.arm.com/documentation/102284/6-16-2LTS/armclang-Reference/Other-Compiler-specific-Features/Half-precision-floating-point-data-types mentions that there are f16 suffixes for literals; that would probably avoid most of these issues. Let me guess; clang doesn't support those literals yet...?

[lntue]

https://developer.arm.com/documentation/102284/6-16-2LTS/armclang-Reference/Other-Compiler-specific-Features/Half-precision-floating-point-data-types mentions that there are f16 suffixes for literals; that would probably avoid most of these issues. Let me guess; clang doesn't support those literals yet...?

@overmighty Do you remember why we didn't use -f16 suffices for literals?

[overmighty]

https://developer.arm.com/documentation/102284/6-16-2LTS/armclang-Reference/Other-Compiler-specific-Features/Half-precision-floating-point-data-types mentions that there are f16 suffixes for literals; that would probably avoid most of these issues. Let me guess; clang doesn't support those literals yet...?

@overmighty Do you remember why we didn't use -f16 suffices for literals?

f16 suffixes were introduced in C++23, but apparently Clang doesn't complain when using them in C++20, even with -pedantic, only GCC does: https://godbolt.org/z/Y4b7adTds.

[nickdesaulniers]

even with -pedantic, only GCC does: https://godbolt.org/z/Y4b7adTds.

Are we compiling with -pedantic? We should not IMO. It's oxymoronic to compile in -std=gnu++XX AND then enable -pedantic. If we wanted to be pedantic, then we could compile in ISO -std=c++XX mode. Rather, I prefer to use the (good) GNU C extensions and avoid being pedantic. We only support clang and gcc, -pedantic helps your code be more portable to other compilers which is a non-goal ATM.