Run LLVM function passes immediately after translating a function to LLVM IR.

[eddyb]

I can see rustc spawning one function pass thread per core, leaving only module passes to account for after trans finishes and the RAM usage peak in trans being lower.
Since we may be translating dependencies first (can someone confirm?), we could even be able to run module passes on isolated function graphs, but that's further into the future.

[thestinger]

There is already a working LLVM changeset by Intel adding support for running function passes in parallel. It will be far slower to create a separate function pass for every function.

[alexcrichton]

Closing, sadly this isn't something that we can do at the frontend level. For components inside of one LLVMContext, you cannot parallelize their usage manually by running pass managers on separate cores. LLVM is built to be parallel at the granularity of the LLVMContext layer right now, but not within (too much shared state).

The real solution for something like this was the changeset that Intel was pushing, but even that was only parallelizing the codegen passes, not the optimization passes (from what I hear). This is certainly something that we need to speed up our builds, but we'll have to put pressure on upstream LLVM (or implement it ourselves) because I don't see any way we could do this manually (without splitting into many LLVMContext-s)