IA64 assembly was "hard" because it wasn't expected to be something that humans would write or debug routinely - it was basically meant for compilers.
This also describes VLIW architectures more generally, and even more so wrt. current speculative developments like Mill, which AIUI doesn't even have a single, standardized "machine-level" assembly; what the machine runs at its lowest level is only defined by your actual chip configuration. (It is understood that GPU compute works much the same way, with the GPU driver acting as a "compiler" of sorts.)