Clang should provide builtin that lowers directly to `llvm.fma.*` intrinsic

[petrhosek]

Clang already provides the __builtin_fma builtin intrinsic which is lowered either to fma call or llvm.fma.* Intrinsic depending on whether math errno is used or not (i.e. -f[no-]math-errno). This behavior makes [__builtin_fma unsuitable for implementing libc fma because unless libc is compiled with -fno-math-errno, on targets where math errno is used, this would result in an infinite recursion.

We encountered this issue in LLVM libc:

llvm-project/libc/src/__support/FPUtil/FMA.h

Lines 27 to 33 in 5eed019

	template <> LIBC_INLINE float fma(float x, float y, float z) {
	return __builtin_fmaf(x, y, z);
	}
	template <> LIBC_INLINE double fma(double x, double y, double z) {
	return __builtin_fma(x, y, z);
	}

On baremetal targets where math errno is used, the __builtin_fma call is lowered to fma call resulting a function trivially calling itself. The current workaround is to compile LLVM libc with -fno-math-errno, but that is generally undesirable.

Ideally, Clang would provide a separate builtin intrinsic that always lowers to llvm.fma.* Intrinsic regardless of math errno behavior. We could then use this intrinsic to implement fma.

[tclin914]

What's expected codegen if that lowers to llvm.fma.* even with -fmath-errno. Does it still emit a function call to fma.

[arsenm]

Can you use __builtin_elementwise_fma for this? I don't think any of the elementwise builtins consider errno