I tested the current state with llvm:min=20 on 151amd64, the same jail that died with 64 assembler errors before. It builds cleanly, 6 packages, none failed. configure picks /usr/local/llvm20/bin/clang and reports "Intel SM4 instructions... yes", so the EVEX path is kept rather than disabled. The solution works. To be honest I would still have preferred the patch. On my 12 core machine with 4 builders llvm20 takes 2:14:33, libgcrypt itself 0:00:46, the whole run 2:16:05, and the llvm20 package is 264 MB. For everyone building from the ports tree this now means a full LLVM toolchain build in order to build a crypto library that is done in 46 seconds. The patch needed four lines and was confirmed by several people on very different hardware. If the USES stays, two remarks on it: sm4-intel-avx512-amd64.S is amd64 only, but the USES applies unconditionally, so aarch64 and i386 build LLVM20 for a problem they do not have. Wrapping the USES in .if ${ARCH} == "amd64" would take care of that. And the commit did not bump PORTREVISION although the package changes.