I can't help but think this should be tracing spans and items

That would be useful for some stuff, but, even ignoring the rustc "timers"...
_{(that we only use an inefficient ad-hoc form of, rn, which is only accessible via RUSTGPU_RUSTFLAGS=-Ztime-passes, while the more advanced measureme APIs are the low-overhead/high-throughput event tracking powering -Zself-profile, which the log/tracing approach can't really ever compete with AFAIK, sadly - but really, few things need that level of profiling, compilers being deterministic and IO-bound is an edge case)}
the most important part here is generating the .spirt/.spirt.html dumps, where the "pass names" show up as table headings - quick example from a dump I have open:

I wouldn't mind automating those strings, but right now they're quite arbitrary, e.g. after Rust-GPU/spirt#25 the mem::analyze_accesses one comes from:

let profiler = before_pass("mem::analyze_accesses", module);
spirt::passes::qptr::analyze_mem_accesses(module, &SPIRT_MEM_LAYOUT_CONFIG);
after_pass(Some(module), profiler);

with spirt::passes::qptr::analyze_mem_accesses defined as:

// FIXME(eddyb) this doesn't really belong in `qptr`.
pub fn analyze_mem_accesses(module: &mut Module, layout_config: &crate::mem::LayoutConfig) {
    crate::mem::analyze::GatherAccesses::new(module.cx(), layout_config)
        .gather_accesses_in_module(module);
}

So there's not even any function named analyze_accesses in a module named mem, at best that could be the case for spirt::passes::mem::analyze_accesses if that function was moved to such a place.

But even if we take the function's path and strip either spirt::passes:: or rustc_codegen_spirv::linker::spirt_passes:: as a prefix (which does sound relatively reasonable, I will admit), how could we obtain that information when that function is in another crate?

You can create trace spans programmatically with custom names, it doesn't have to be a function nor a proc macro. Tracing is in rustc proper and lots of crates (not sure what is actually instrumented though).

But yeah, it doesn't help with table headings unless you installed your own subscriber and collected and then output (that might actually be interesting).