Email or username:

Password:

Forgot your password?
notgull's posts Post Back to profile
Top-level
notgull

@cr1901 My plan is to replace it with the Cranelift backend. It's written entirely in Rust, so it should be usable if we have a Rust compiler.

If that doesn't work, my plan is to write a QBE backend for rustc and use that instead.

25 August at 21:20 | Open on hachyderm.io
8 comments
William D. Jones

@notgull Does Cranelift do optimizations? I've heard conflicting info about this ("it doesn't optimize", "it does optimize, but debug-friendly opts only, no further opts planned", "release opts are planned/in progress").

25 August at 21:21 | Open on mastodon.social
notgull

@cr1901 I don't know, as far as I know there are some optimization passes but it's nowhere near LLVM's caliber.

25 August at 21:22 | Open on hachyderm.io
William D. Jones

@notgull Actual no-optimization builds of rustc are so slow that config.toml itself says they're unsupported. Hopefully getting an acceptable speed compiler doesn't require LLVM-like levels of effort (i.e. 80% of the speed for 20% of the work).

Anyways, if you succeed in this project, I will have to find something else to bit- err, nothing left to bitch about- wrt rustc's bootstrap situation, so I do hope its feasible, and that e.g. upstream (if QBE is required) and Cranelift are supportive :D.

25 August at 21:28 | Open on mastodon.social
William D. Jones

@notgull This is OT, but while its on my mind...

Is GH Discussions a reasonable place to ask questions about design decisions about smol? (Specifically, I have a lifetime question, and can't figure it out for the life of me)

29 August at 18:39 | Open on mastodon.social
bjorn3

@notgull @cr1901 There are indeed a couple of optimizations, but not much. A cg_llvm compiled cg_clif builds the standard library in 19s. A cg_clif compiled rustc/cg_clif builds the standard library in 4m 42s, or in other words about a 14x slowdown. github.com/rust-lang/rustc_cod To shrink the gap, implementing clif ir level inlining (which should be more effective than mir inlining) as well as better optimizations of loads and stores should help a fair bit.

@notgull @cr1901 There are indeed a couple of optimizations, but not much. A cg_llvm compiled cg_clif builds the standard library in 19s. A cg_clif compiled rustc/cg_clif builds the standard library in 4m 42s, or in other words about a 14x slowdown. github.com/rust-lang/rustc_cod To shrink the gap, implementing clif ir level inlining (which should be more effective than mir inlining) as well as better optimizations of loads and stores...

25 August at 21:52 | Open on hachyderm.io
William D. Jones

@bjorn3 @notgull Idk about 14x, but I think even a 3x slowdown would be acceptable to me personally :P.

25 August at 22:00 | Open on mastodon.social
Go Up