Stream: cranelift

Topic: Generate 128bit IDIV

view this post on Zulip Andrix (Sep 18 2024 at 12:07):

Hello, I'm trying to generate an x64 IDIV instruction. I have 2 64bit input values and a 64bit divisor. I tried to iconcat the inputs and sextend the divisor to 128 bit values and then use sdiv and srem to calculate the results. The problem is that 128bit division seems unimplemented and isle.rs/put_in_reg panics when trying to assign a 128bit value to a 64bit register. Is there any way to do this cleanly? Or do I have to manually calculate everything? Would that then simplify to a single x64 IDIV in the end?

view this post on Zulip Notification Bot (Sep 18 2024 at 14:57):

This topic was moved here from #general > Generate 128bit IDIV by fitzgen (he/him).

view this post on Zulip Alex Crichton (Sep 18 2024 at 15:01):

I believe you're right that Cranelift doesn't have support for this. Other compilers I've seen implemente 128-bit division with a libcall (e.g. __udivti3) and that would most likely be Cranelift's own fallback libcall as well.

I don't think that x64 idiv matches CLIF semantics though because I believe the instruction itself semantically divides a 128-bit integer by a 64-bit integer producing a 64-bit division result and a 64-bit remainer. That doesn't cleanly map to CLIF sdiv or udiv because you'd also need an ireduce on the outside explicitly discarding the upper bits

view this post on Zulip fitzgen (he/him) (Sep 18 2024 at 15:02):

we could consider adding an ISLE rule to match that case, fwiw

view this post on Zulip Alex Crichton (Sep 18 2024 at 15:04):

I think it'd look something like

(has_type $I64 (ireduce (sdiv x (has_type $I64 (sextend y)))))

for a pattern here (which may or may not be what you're looking for

view this post on Zulip Alex Crichton (Sep 18 2024 at 15:04):

it may also require edits to the MInst enum to support the full generality here, I haven't looked at div myself in a while

view this post on Zulip Andrix (Sep 20 2024 at 10:12):

So it's not really possible at the moment without modifying Cranelift, right? Maybe I'll try to insert it somehow after Cranelift has finished running.

view this post on Zulip Alex Crichton (Sep 20 2024 at 16:52):

Right yeah, there's no lowering rules in Cranelift where the upper bits of the division operation are customized. They're always zero or sign-extended from the lower 64-bits

Last updated: Dec 23 2024 at 12:05 UTC