ld: error: undefined reference: roundevenf

[brad0]

OS: OpenBSD
Archs: powerpc, sparc64, etc.
Compiler in use on most archs Clang (19), though GCC (15) on sparc64.

I noticed VVenC failing to build on a few archs with an error as shown in the title. I didn't see why at first as I did not see any references to the roundevenf function in the VVenC code base, but then remembered there is a sub-compontent which github search wouldn't look through and
sure enough with the tarball it was there and came across a bunch of references to roundevenf.

I noticed another issue with the same error, but though I should keep this as a separate issue.

http://build-failures.rhaalovely.net/powerpc/2026-01-28/multimedia/vvenc.log

[mr-c]

Hello @brad0 ; this seems like a clang issue ; perhaps your clang is not built/configured to use the compilter-rt library which is responsible for providing roundevenf. For example, on Debian we had to enable compiler-rt for s390x: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=1114980 https://salsa.debian.org/pkg-llvm-team/llvm-toolchain/-/commit/1d5b5e7681ee80cb643b57396ef54cf3d458f5a4

[mr-c]

While you wait for that to be fixed, you can locally patch

simde/simde/simde-math.h

Lines 1268 to 1306 in c3e0eb1

	#if !defined(simde_math_roundeven)
	#if \
	((!defined(HEDLEY_EMSCRIPTEN_VERSION) || HEDLEY_EMSCRIPTEN_VERSION_CHECK(3, 1, 43)) && HEDLEY_HAS_BUILTIN(__builtin_roundeven)) || \
	HEDLEY_GCC_VERSION_CHECK(10,0,0)
	#define simde_math_roundeven(v) __builtin_roundeven(v)
	#elif defined(simde_math_round) && defined(simde_math_fabs)
	static HEDLEY_INLINE
	double
	simde_math_roundeven(double v) {
	double rounded = simde_math_round(v);
	double diff = rounded - v;
	if (HEDLEY_UNLIKELY(simde_math_fabs(diff) == 0.5) && (HEDLEY_STATIC_CAST(int64_t, rounded) & 1)) {
	rounded = v - diff;
	}
	return rounded;
	}
	#define simde_math_roundeven simde_math_roundeven
	#endif
	#endif
	#if !defined(simde_math_roundevenf)
	#if \
	((!defined(HEDLEY_EMSCRIPTEN_VERSION) || HEDLEY_EMSCRIPTEN_VERSION_CHECK(3, 1, 43)) && HEDLEY_HAS_BUILTIN(__builtin_roundevenf)) || \
	HEDLEY_GCC_VERSION_CHECK(10,0,0)
	#define simde_math_roundevenf(v) __builtin_roundevenf(v)
	#elif defined(simde_math_roundf) && defined(simde_math_fabsf)
	static HEDLEY_INLINE
	float
	simde_math_roundevenf(float v) {
	float rounded = simde_math_roundf(v);
	float diff = rounded - v;
	if (HEDLEY_UNLIKELY(simde_math_fabsf(diff) == 0.5f) && (HEDLEY_STATIC_CAST(int32_t, rounded) & 1)) {
	rounded = v - diff;
	}
	return rounded;
	}
	#define simde_math_roundevenf simde_math_roundevenf
	#endif
	#endif

to skip the builtins on the affected architectures and use the slower fallbacks.

[brad0]

Hello @brad0 ; this seems like a clang issue ; perhaps your clang is not built/configured to use the compilter-rt library which is responsible for providing roundevenf. For example, on Debian we had to enable compiler-rt for s390x: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=1114980 https://salsa.debian.org/pkg-llvm-team/llvm-toolchain/-/commit/1d5b5e7681ee80cb643b57396ef54cf3d458f5a4

I don't see the functions (roundeven / roundevenf) as part of compiler-rt. All of our Clang archs (powerpc) uses the compiler-rt runtime. The only place I have run into roundeven / roundevenf is within glibc's libm implementation. I noticed an issue about musl users complaining. None of the BSD's have these libm functions. These functions seem to be added in C23 which is extremely new.

[mr-c]

@brad0 Yeah, I agree that it is very messy that clang defines builtins that are not available in its own runtime. I guess we could also merge a fix locally, to skip the builtin for clang on certain archs.

[brad0]

@brad0 Yeah, I agree that it is very messy that clang defines builtins that are not available in its own runtime. I guess we could also merge a fix locally, to skip the builtin for clang on certain archs.

Not fully understanding the code in question, but also seeing that builtins can be lowered to a lib call function I was looking at Clang and trying to test whether the builtin was lowered to a function but I didn't see any indication that that was the case. If I saw any indication that was the case I would submit a patch to fix it.

[brad0]

Doing a bit more testing I see why I didn't see this initially. If I target my test with the builtin to aarch64 it lowers to an instruction. If I target any other arch, x86_64, i386, powerpc, powerpc64, sparc64 it lowers to a libcall.

[brad0]

Also to add to this this isn't just a Clang issue as we ran into the same error on sparc64 which uses GCC 15.

[brad0]

6b772fe Fixes the issue if using GCC. But the same issue affects the use of the builtin if using Clang. This is also the same issue on the FreeBSD issue here #1014. This needs to be re-opened until Clang is fixed.