Recompiling a program for a different instruction set architecture without access to its source is a long-standing goal of binary analysis, and one that repeatedly runs aground on details the source language never exposed. We report on extending a dynamic binary lifting framework to translate x86 binaries to AArch64 by way of LLVM IR, and on the architectural mismatches that make the translation hard in practice — calling-convention divergence, the absence of size information for pointer arguments at library boundaries, and the gap between recorded dynamic execution and the paths a recompiled binary may actually take. We validate the resulting pipeline on real-world SPEC CPU 2017 programs.