You've mentioned Intel's compilers a few times but they're not a magic thing that could make Itanic not-suck. Suppose they make general purpose code that's twice as fast as GCC when the whole computer is dedicated to running one specific process. They don't, but let's pretend. Even in that ideal scenario, it falls apart as soon as you're processing unpredictable input. VLIW sucks hard at chewing through data that's not precisely what it's expecting. You could probably make a video codec that performed really well, but it would be impossible to make a database that performed consistently not-badly. It's not that the existing compilers weren't good enough, not even the magic Intel ones. It's that it's not possible to pre-generate VLIW opcodes that do well when the input value isn't a homogeneous stream. Even if Intel's code was 2x better than GCC's — and again, it wasn't — twice abominable was still abominable.

I wasted more time SSHed into an Itanic than most people had to, so I'm speaking from first-hand experience. Some very, very specific Itanium code was a bit faster than the equivalent x86 or similar. That was always some very tightly scoped thing that did the exact same tight loop a gazillion times, like en-/decrypting a data stream. Anything more heterogenous ran poorly, as in multiple times the wall-clock time of the same workload on the x86 server next to it that cost a tenth as much. And that's even when using the magic Intel compilers.

Itanium was bad. The compiler tech didn't exist to make it not-bad, and in retrospect I think it become obvious that it couldn't exist. It wasn't a matter of the compiler authors needing to be clever. It was more like making Itanium live up to the hype required making P=NP.