Alternative regalloc backends
Amanieu opened this issue · 2 comments
The regalloc2 crate provides a good API (VReg, PReg, Operand, Function, etc) for interacting with a register allocator. It would be nice if other register allocator backends could be made available with the same API to simplify integration on the embedder side.
What I specifically have in mind is a "fast" backend that simply allocates registers in a single pass with the goal of minimizing compilation time. This will have a huge impact on compilation time since my compiler spends 80% of its time in register allocation. In practice I expect to use this for a tiering JIT where hot code can later be re-compiled using the slow regalloc backend that produces better code.
Agreed -- either a linear-scan (in the classical sense, with precomputed live ranges and evicting furthest-future-used values) or a true single-pass allocator (as baseline compilers typically do, evicting all values to stackslots at merge-points and evicting all or per LRU on running out of registers) would be great to have. I don't have the bandwidth to work on this now but I'd be happy to review it if you or someone else wants to work on this.
It's not a high priority for me at the moment, but I'm glad to hear that you're happy to have it as part of the regalloc2 crate.