Stream: cranelift

Topic: Interleaving `Inst` and `Value` nodes?

Bongjun Jang (Oct 19 2025 at 10:55):

Hi, I'm interested in writing an optimization (((X / Y) = N) = ((Y * N) = X)) in ISLE.
The left-hand side will have something like (icmp (IntCC.Equal) (udiv x y) ...).
However, the problem is that the udiv term yields an Inst node while the icmp term requires Value nodes as operands.
It makes writing such a rule very difficult.

Is there any walkaround possible in ISLE? Thank you!

Chris Fallin (Oct 20 2025 at 00:14):

It looks like there is some glue missing here:

• In the lowering prelude of ISLE (i.e., a different set of definitions/bindings used when compiling machine backends, rather than mid-end opt rules), we have an auto-converter defined here.
• We define the corresponding Rust helper here, i.e. present for both mid-end and lowering environments.
• We define the actual FFI signature here, i.e., also common to both environments.

• So it appears that the only missing piece is moving the

  (convert Inst Value def_inst)

  line from the lowering-specific prelude to the common prelude. I just checked that doing so causes a simple rule of the form you show to typecheck in the mid-end. (I didn't take the time to come up with a valid rule and check its full execution but please let me know if that doesn't work!)

Chris Fallin (Oct 20 2025 at 00:15):

Feel free to do so as part of an upcoming PR if you have one, or separately and tag me for review, either way.

Chris Fallin (Oct 20 2025 at 00:17):

(and, goes without saying but: thank you for continuing to work on optimizations of this sort! the slow but steady improvements to our codegen are greatly appreciated)

Bongjun Jang (Oct 20 2025 at 00:52):

Thank you for the detailed answer! I'm trying to generate valid ISLE rules as many as possible.

For that you mentioned improvements for cranelift's codegen,
I wonder if we can precisely measure the improvements on generated code, and the compilation overhead. At the moment, I can think of running cranelift on some benchmarks and measure the time spent. Is there any established method for Cranelift per se?

Chris Fallin (Oct 20 2025 at 04:14):

We normally use Sightglass for that; see this part of the readme in particular showing how to build two different .sos from Wasmtime and compare them

Chris Fallin (Oct 20 2025 at 04:15):

the benchmark suite still isn't super-well filled out, but at least will give you something. Typically the numbers that are most interesting to the core Cranelift folks are the non-microbenchmarks, which more or less is the default suite plus maybe regex and blake3-*

Bongjun Jang (Oct 20 2025 at 04:38):

Thank you very much. I hope the improved codegen can make some change. I'll try the benchamrks.

Last updated: Dec 06 2025 at 07:03 UTC