While they used Cranelift IR itself (amongst others, not just LLVM) to show performance improvements (thus making it complementary and not a replacement) you raise a good point. Quite possible it is not as full-featured yet so perhaps in the future, if at all.
The TPDE-based back-end compiles 4.27x faster than Cranelift and 2.68x faster than Cranelift with its fast register allocator, but is 1.74x slower than Winch
They're hitting another design point on the compile time vs. code-quality tradeoff curve, which is interesting. They compile 4.27x faster than Cranelift with default (higher quality) regalloc, but Cranelift produces code that runs 1.64x faster (section 6.2.2).
This isn't too surprising to me, as the person who wrote Cranelift's current regalloc (hi!) -- regalloc is super important to run-time perf, so for Wasmtime's use-case at least, we've judged that it's worth the compile time.
TPDE is pretty cool and it's great to see more exploration in compiler architectures!
The TPDE-based back-end compiles 4.27x faster than Cranelift and 2.68x faster than Cranelift with its fast register allocator, but is 1.74x slower than Winch